Let a host marked for the relay actually be reached through it

The relay checkbox said it was not wired up, and the plan said the server
half was built and shipped. It was not: there was ADR 0004, the contract
records, RelayOptions and a CHECK constraint, and RelayOptions said in its
own remarks that there was no endpoint and no dialler. So both halves are
built here, as step 3 of docs/reaching-a-host-you-cannot-dial.md.

Server, in Features/Relay:

- POST /api/v1/relay/tickets makes every decision: the host, Connect on
  its vault, the relay switched on, and the stored name resolved once and
  judged by RelayAddressPolicy. Loopback, link-local, 0.0.0.0/8,
  multicast and the metadata endpoints are refused with no configuration
  escape, with IPv4-mapped and NAT64 spellings unwrapped first; then the
  operator's denials, allow-list and private-network switch. What passes
  goes into a Data Protection ticket, and nothing resolves the name again.
  Every target refusal is the same 403, so the endpoint cannot be used to
  map the operator's DNS; the real reason is logged.
- GET /api/v1/relay/connect takes the ticket as a second subprotocol,
  checks its expiry against TimeProvider and spends it with a unique
  insert into relay_ticket_use before accepting the upgrade. After the
  upgrade a refusal is a close code: 4502 for a host the server could not
  reach, 4429 for a limit, 4408 idle, 4409 lifetime, 1001 after a drain.
- relay_session is the audit row, written from short-lived scopes.

Client:

- RelayBridge is a loopback SOCKS5 listener that accepts once and pipes
  to the relay socket. It rides on the new SshLoopbackProxy.Lifetime, so
  the connection disposes it on every way a session ends and a failed
  connect disposes it before throwing.
- ServerRelay, exposed as IVaultServer.Relay from /api/v1/meta, asks for
  the ticket, opens the socket, starts the bridge and turns every refusal
  into a sentence.
- VaultViewModel and TransfersViewModel reach a relayed host through it,
  never fall back to dialling it directly, and replace SSH.NET's transport
  error with the relay's own reason when it gave one.

Neither side ever cancels a WebSocket operation. Cancelling one aborts
the socket, and an aborted socket sends no close frame, which is the only
way the client learns the server could not reach the host.
AnIdleSession_IsClosedWithTheIdleCode asserts that the code arrives.

Three departures from ADR 0004, recorded in it: awaited copy loops rather
than Pipelines (same backpressure, two 32 KiB buffers), no ?ticket=
fallback, and no device id in the ticket.

Tests: API 245, shell 582 (5 skipped), layout 170, SSH 110 including a
real login through the bridge, API client 30, contracts 25. Not run: the
Android build (its change is copy only) and the Alpine layout oracle.

.gitattributes

@@ -20,8 +20,14 @@
 *.sql       text
 
 # Files that must keep LF regardless of platform — shell scripts run in Linux
-# containers, and Dockerfiles are read by a Linux daemon.
+# containers, Dockerfiles are read by a Linux daemon, and an RPM spec is a shell
+# script in disguise: rpmbuild feeds its %install section to /bin/sh, where a
+# carriage return at the end of a line is an argument rather than nothing. A udev
+# rule is parsed by udev's own lexer, which treats a stray carriage return as part
+# of the value it is reading.
 *.sh          text eol=lf
+*.spec        text eol=lf
+*.rules       text eol=lf
 Dockerfile    text eol=lf
 *.Dockerfile  text eol=lf
 Caddyfile     text eol=lf

.github/workflows/ci.yml

@@ -29,6 +29,18 @@ jobs:
   build:
     name: build and test
     runs-on: [linux]
+    # ◆ A CEILING, BECAUSE THIS FILE HAD NONE AND A HUNG TEST RAN UNTIL SOMEBODY NOTICED.
+    #
+    # One test in DodoSSH.Client.App.Tests waited on a named pipe that, on Linux, nothing was ever going
+    # to answer — and it waited on a token with no timeout, so the job did not fail, it sat there. It was
+    # killed by hand after twenty minutes, having reported nothing but the eighteen suites that had
+    # already passed.
+    #
+    # Twenty-five minutes against a job that ordinarily takes three. It is not a performance budget and
+    # must not be read as one: a build that legitimately grows past this should have the number raised.
+    # What it buys is that the failure mode is a red job with a named step rather than a runner somebody
+    # has to go and stop.
+    timeout-minutes: 25
     steps:
       # act_runner runs every `uses:` action with node inside the job container, and this
       # runner's image has none — the run died on the first line of actions/checkout with
@@ -220,7 +232,12 @@ jobs:
       - name: build
         run: dotnet build DodoSSH.slnx --no-restore --configuration Release
 
+      # Tighter than the job's own ceiling, and deliberately so: a hang here is the likeliest of the
+      # lot — it is the only step running code written for this repository rather than a tool — and a
+      # step-level timeout names the step in the failure instead of leaving a reader to work out which
+      # of fifteen was still running. Ten minutes against a suite that takes two.
       - name: test
+        timeout-minutes: 10
         run: dotnet test DodoSSH.slnx --no-build --configuration Release
 
       # The one build shape nothing else here exercises: a self-contained RID-specific publish. Its
@@ -291,6 +308,7 @@ jobs:
       # so it is not done either. What reaches users is built, installed and walked through Phase 16
       # of docs/manual-checks.md by a person first.
       - name: package the windows desktop client
+        id: winpack
         if: github.event_name != 'pull_request'
         run: |
           set -euo pipefail
@@ -374,6 +392,142 @@ jobs:
           ls -la "$releases"
           echo "Packaged DodoSSH $packVersion for win-x64, from a build MinVer calls $version."
 
+          # Handed to the macOS step below rather than worked out again there. The floor logic above
+          # is thirty lines of reasoning about MinVer's pre-first-tag answer, and a second copy of it
+          # is a second thing to keep in step — while two desktop packages built from one commit
+          # carrying different version numbers is precisely the confusion this file spends that
+          # reasoning to avoid.
+          echo "packVersion=$packVersion" >> "$GITHUB_OUTPUT"
+
+      # ◆ AND THE macOS BUNDLE IS BUILT HERE, ON LINUX, AND IS ALSO THROWN AWAY.
+      #
+      # Same argument as the Windows step above, one platform along: the failures a release is most
+      # exposed to are the ones only the packager finds, and the person who would otherwise find them
+      # is the one midway through a release on the one Mac that can cut one.
+      #
+      # What this catches that the Windows step cannot: the osx-arm64 restore graph. A native package
+      # that resolves for win-x64 and has no osx-arm64 asset — libsodium and SkiaSharp both ship per
+      # RID — fails here, on every main build, rather than at the first `dotnet publish` of a release
+      # nobody can retry without a Mac.
+      #
+      # ◆ bundle, NOT pack, AND THE DIFFERENCE IS NOT A CHOICE.
+      #
+      # `vpk [osx]` cross-compiling from a non-Mac offers exactly one packaging verb: bundle, which
+      # builds the .app. There is no `[osx] pack` off a Mac, and that is correct rather than a gap —
+      # pack signs with codesign, submits to Apple with notarytool and staples the ticket, all of
+      # which is Apple tooling that exists on no other platform. So this proves the bundle and stops
+      # where the platform does.
+      #
+      # No --plist and no --icon either, deliberately. Both are proved by scripts/release-macos.sh on
+      # the machine that can also check the result; passing a rendered plist here would mean copying
+      # the substitution out of that script to no end, since nothing looks at what this produces.
+      #
+      # ◆ NOTHING IS UPLOADED, FOR THE REASON THE WINDOWS STEP GIVES.
+      #
+      # RUNNER_TEMP, dying with the job. ADR 0013 rule 3 puts the capability to ship somebody a build
+      # on a machine which is not a runner, and an unsigned .app is additionally something no Mac
+      # would open — so publishing it would be handing out a file whose only possible use is confusion.
+      - name: publish and bundle the macos desktop client
+        if: github.event_name != 'pull_request'
+        run: |
+          set -euo pipefail
+
+          # RestoreLockedMode=false for the RID, exactly as the win-x64 publish above does — see the
+          # long note there for why the committed lock files are deliberately RID-free. This runner's
+          # checkout is thrown away, so the lock files it rewrites go nowhere.
+          dotnet publish src/DodoSSH.Client.App/DodoSSH.Client.App.csproj \
+            --configuration Release --runtime osx-arm64 --self-contained true \
+            -p:RestoreLockedMode=false \
+            --output "$RUNNER_TEMP/osx-arm64"
+
+          # The apphost has no extension on macOS, so this is `DodoSSH` and not `DodoSSH.exe`. Named
+          # rather than globbed, because a publish that produced no apphost at all would otherwise
+          # bundle happily and produce an .app that launches nothing.
+          if [ ! -s "$RUNNER_TEMP/osx-arm64/DodoSSH" ]; then
+            echo "The osx-arm64 publish produced no apphost." >&2
+            ls -la "$RUNNER_TEMP/osx-arm64" >&2 || true
+            exit 1
+          fi
+
+          bundles="$RUNNER_TEMP/osx-bundle"
+
+          # The quotes around [osx] are load-bearing, exactly as they are on '[win]' above: unquoted,
+          # the shell reads it as a glob matching any one of o, s and x.
+          dotnet vpk '[osx]' bundle \
+            --skip-updates \
+            --packId DodoSSH.Desktop \
+            --packVersion '${{ steps.winpack.outputs.packVersion }}' \
+            --packDir "$RUNNER_TEMP/osx-arm64" \
+            --packTitle DodoSSH \
+            --packAuthors DodoTech \
+            --mainExe DodoSSH \
+            --bundleId dev.dodotech.dodossh \
+            --runtime osx-arm64 \
+            --channel osx \
+            --outputDir "$bundles"
+
+          # Asked for rather than inferred from an exit code, for the reason the Windows step gives.
+          # The Info.plist is the specific thing worth naming: a bundle missing it is a directory
+          # macOS will not treat as an application at all, and it is the one part of the .app that
+          # vpk composes rather than copies.
+          app="$bundles/DodoSSH.Desktop.app"
+          if [ ! -s "$app/Contents/Info.plist" ]; then
+            echo "vpk reported success and there is no Info.plist at $app/Contents/Info.plist." >&2
+            find "$bundles" -maxdepth 3 >&2 || true
+            exit 1
+          fi
+
+          echo "Bundled DodoSSH ${{ steps.winpack.outputs.packVersion }} for osx-arm64."
+
+      # ◆ AND THE THIRD DESKTOP PLATFORM, WHICH IS THE ONE THIS RUNNER ACTUALLY IS.
+      #
+      # Same argument as the two steps above and with none of their cross-compiling: this job runs on
+      # Linux, so the linux-x64 publish here is a native one. What it proves that the others cannot is
+      # the linux-x64 restore graph — libsodium, SkiaSharp and HarfBuzzSharp all ship native assets per
+      # RID, and a package that resolves for win-x64 and has no linux-x64 asset fails here.
+      #
+      # The runner is still Alpine and the package targets glibc distributions, which is fine and is not
+      # glossed over: scripts/ci-linux-rpm.sh explains why rpmbuild composing an RPM for one libc on a
+      # host running another is correct rather than lucky.
+      #
+      # ◆ IT IS ALSO WHERE THE PACKAGING ITSELF IS PROVED, WHICH ON THIS PLATFORM IS NOT A FORMALITY.
+      #
+      # The Windows and macOS steps hand a directory to vpk and vpk decides what the package needs.
+      # There is no vpk here — Velopack's Linux target is an AppImage and this is a distribution
+      # package, updated by dnf rather than by the application — so the dependency metadata is
+      # generated by rpm from the published tree, and scripts/ci-linux-rpm.sh interrogates the result
+      # rather than trusting an exit code. Every check in there is a way a package can be built
+      # successfully and then refuse to install on every host in the world; one of them was written
+      # because .NET's own libmscordbi.so links libmscordaccore.so, which is exactly that.
+      #
+      # Release-shaped rather than nightly-shaped — package name dodossh, tree at /opt/DodoSSH — so
+      # that what is proved on every main build is the packaging the release channel would use. The
+      # desktop job below builds the nightly-shaped one and is the only job that publishes either.
+      #
+      # ◆ NOTHING IS UPLOADED, FOR THE REASON THE WINDOWS STEP GIVES.
+      #
+      # RUNNER_TEMP, dying with the job. An RPM is if anything the easiest of the three to pass on
+      # casually — one file, no installer to click through, `dnf install` and done — which is a reason
+      # to keep it off a downloads page rather than a reason to put it on one. ADR 0013 rule 3.
+      - name: publish and package the linux desktop client
+        if: github.event_name != 'pull_request'
+        run: |
+          set -euo pipefail
+
+          # RestoreLockedMode=false for the RID, exactly as the two publishes above do — see the long
+          # note on the win-x64 one. This checkout is thrown away, so the lock files it rewrites go
+          # nowhere.
+          dotnet publish src/DodoSSH.Client.App/DodoSSH.Client.App.csproj \
+            --configuration Release --runtime linux-x64 --self-contained true \
+            -p:RestoreLockedMode=false \
+            --output "$RUNNER_TEMP/linux-x64"
+
+          scripts/ci-linux-rpm.sh \
+            --payload "$RUNNER_TEMP/linux-x64" \
+            --version '${{ steps.winpack.outputs.packVersion }}' \
+            --channel release \
+            --out "$RUNNER_TEMP/linux-rpm"
+
       # This includes the end-to-end suite, which starts PostgreSQL, Keycloak and an OpenSSH
       # server through Testcontainers and runs the API as a child process — so it needs a
       # Docker daemon and gets one here. That is why the tests run on ubuntu rather than
@@ -419,6 +573,8 @@ jobs:
           exit 0
   android:
     name: android head
+    # See the build job. The docker pull and the toolchain image make this one legitimately slower.
+    timeout-minutes: 40
     runs-on: [linux]
     # Writes the nightly release at the end of the job; see that step for what the capability is and why
     # it is acceptable here and nowhere else. Job-scoped, so nothing else in this file gains it.
@@ -778,7 +934,10 @@ jobs:
           # never made.
           published="$(curl -fsS -H "$auth" "$api/releases/$release")"
           for name in $names; do
-            if ! printf '%s' "$published" | grep -qF "\"name\":\"$name\""; then
+            # grep without -q, which has the same exit status and reads to the end of its input. A
+            # reader that stops early under pipefail turns the writer's SIGPIPE into this condition's
+            # answer; see the note in scripts/ci-linux-rpm.sh, where that cost a green build.
+            if ! printf '%s' "$published" | grep -F "\"name\":\"$name\"" >/dev/null; then
               echo "The release was created but $name is not on it:" >&2
               printf '%s\n' "$published" >&2
               exit 1
@@ -787,8 +946,380 @@ jobs:
 
           echo "Published nightly $VERSION with$names"
 
+  desktop:
+    name: desktop nightly
+    # See the build job. Two publishes, a pack and an upload of about 120 MB.
+    timeout-minutes: 40
+    # Gated on the tests, like the image job and for a stronger version of its reason: this one is
+    # installed by people and replaces itself afterwards. Nothing anybody runs should come out of a
+    # commit whose suite was red.
+    needs: [build]
+    runs-on: [linux]
+    # main only, and the whole job rather than its last step. A v* tag belongs to the release channel,
+    # which no runner may publish — ADR 0013 rule 3 — and the desktop head is already built and packed
+    # on tags by the build job above, so there is nothing here a tag build would gain.
+    if: github.ref == 'refs/heads/main'
+    # Writes the rolling nightly release at the end of the job. Job-scoped, so no other job in this file
+    # gains it; see the publish step for what the capability is and why this channel may hold it.
+    permissions:
+      contents: write
+    steps:
+      # Duplicated from the build job; see the comment there for why it cannot be factored out. There
+      # are four copies now, and any change has to be made in all four.
+      - name: ensure node and git
+        run: |
+          set -eu
+          SUDO=""
+          [ "$(id -u)" -eq 0 ] || SUDO="sudo"
+
+          missing=""
+          command -v node >/dev/null 2>&1 || missing="$missing nodejs"
+          command -v git  >/dev/null 2>&1 || missing="$missing git"
+
+          if [ -z "$missing" ]; then
+            echo "node $(node --version), git $(git --version)"
+            exit 0
+          fi
+
+          echo "Installing:$missing"
+          if command -v apt-get >/dev/null 2>&1; then
+            $SUDO apt-get update -qq
+            $SUDO apt-get install -y --no-install-recommends $missing
+          elif command -v apk >/dev/null 2>&1; then
+            $SUDO apk add --no-cache $missing
+          elif command -v dnf >/dev/null 2>&1; then
+            $SUDO dnf install -y $missing
+          else
+            echo "No apt-get, apk or dnf here, so node cannot be installed from inside the" >&2
+            echo "job. Point the runner's container.image at something that ships node." >&2
+            exit 1
+          fi
+          echo "node $(node --version), git $(git --version)"
+
+          major="$(node --version | sed 's/^v//; s/\..*//')"
+          if [ "$major" -lt 20 ]; then
+            echo "::warning::node $major is older than the runtime these actions target;" \
+                 "give the runner an image with node 20 or newer if actions misbehave."
+          fi
+
+      # fetch-depth 0 for the reason the other three jobs give, and here it decides what gets published:
+      # MinVer's answer is this build's version and the number a nightly client compares against.
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          fetch-depth: 0
+
+      - uses: actions/setup-dotnet@a98b56852c35b8e3190ac28c8c2271da59106c68 # v6.0.0
+        with:
+          global-json-file: global.json
+          cache: true
+          cache-dependency-path: '**/packages.lock.json'
+
+      # Before the version is read, and that ordering is not tidiness. MinVer arrives as a package, so its
+      # target does not exist until a restore has written obj/*.nuget.g.targets — and `-t:MinVer` on an
+      # unrestored project fails MSB4057 "the target does not exist", which reads like a typo in this file
+      # rather than like a missing restore. The build job's tag check is safe because it runs after that
+      # job's own restore; this job has none, so it needs this one.
+      #
+      # Locked, like the solution restore in the build job. The RID-specific restore the publish needs is
+      # unlocked and asks for that itself, exactly as the build job's publish does.
+      - name: restore
+        run: dotnet restore src/DodoSSH.Client.App/DodoSSH.Client.App.csproj --locked-mode
+
+      # ◆ THE VERSION IS DECIDED ONCE HERE AND THEN FORCED ON EVERYTHING.
+      #
+      # MinVer's own answer is not usable as it stands: until a v* tag exists it is 0.0.0-alpha.0.N, and
+      # vpk refuses to pack anything below 0.0.1. The floor is applied to the *whole build* rather than to
+      # the packaging alone, through MinVerVersionOverride, and that is the part worth understanding.
+      #
+      # Packing a version the assemblies disagree with would put one number in the installer and another
+      # on the preferences screen — the screen a person reads when asked which nightly they are on, and
+      # the number they would then quote into an issue that nobody can match to a build. MinVer sets both
+      # Version and InformationalVersion from the override, so the two cannot drift.
+      #
+      # Monotonic across the boundary, which is what a feed needs: heights keep rising within a floored
+      # version, and the first real tag moves the whole number up past every floored one.
+      - name: the nightly version
+        id: version
+        run: |
+          set -euo pipefail
+
+          # -t:MinVer for the reason the tag check in the build job spells out: without a target named,
+          # -getProperty answers the SDK default and every nightly would claim the same version forever.
+          version="$(dotnet msbuild src/DodoSSH.Client.App/DodoSSH.Client.App.csproj \
+                       -getProperty:Version -t:MinVer -nologo | tr -d '[:space:]')"
+
+          if [ -z "$version" ]; then
+            echo "Could not read the version from MSBuild." >&2
+            exit 1
+          fi
+
+          case "$version" in
+            0.0.0*)
+              floored="0.0.1${version#0.0.0}"
+              echo "MinVer says $version, which vpk will not pack; this nightly is $floored."
+              version="$floored"
+              ;;
+          esac
+
+          echo "version=$version" >> "$GITHUB_OUTPUT"
+          echo "$version"
+
+      # DodoChannel=nightly is what makes this a different application rather than the same one built
+      # again: it puts the channel in the assembly, which is where DesktopChannel reads it to pick the
+      # feed to poll, whether to accept prereleases, and which profile directory to keep a cache in.
+      #
+      # RestoreLockedMode=false for the RID, exactly as the build job's publish does — see the long note
+      # there. This checkout is thrown away, so the lock files it rewrites go nowhere.
+      - name: publish the nightly
+        env:
+          VERSION: ${{ steps.version.outputs.version }}
+        run: >
+          dotnet publish src/DodoSSH.Client.App/DodoSSH.Client.App.csproj
+          --configuration Release --runtime win-x64 --self-contained true
+          -p:RestoreLockedMode=false
+          -p:DodoChannel=nightly
+          -p:MinVerVersionOverride="$VERSION"
+          --output "$RUNNER_TEMP/nightly-win-x64"
+
+      # ◆ A DIFFERENT PACK ID, A DIFFERENT CHANNEL, A DIFFERENT TITLE. ALL THREE, AND NONE IS COSMETIC.
+      #
+      # packId decides where Velopack installs and what an installed client matches an update against, so
+      # DodoSSH.Desktop.Nightly is what makes this install *beside* the release build rather than over it,
+      # and what stops either feed's package being applied to the other's install.
+      #
+      # channel decides the name of the index file on the feed — releases.win-nightly.json — and it is a
+      # contract with VelopackUpdateChannel.NightlyChannel. Disagree on this word and the channel answers
+      # nothing, forever, with no error anywhere.
+      #
+      # title is what a person reads in the Start menu and in Add/Remove Programs, and it is the only one
+      # of the three they will ever see. Two entries both called DodoSSH would be the whole benefit of
+      # installing side by side, thrown away at the last step.
+      #
+      # No `vpk download` and so no deltas: this channel deletes its previous release on every push, so
+      # there would be nothing on the feed for a delta to be applied against. A nightly update is a full
+      # download, which is the honest cost of a rolling channel that keeps exactly one build.
+      - name: package the nightly
+        env:
+          VERSION: ${{ steps.version.outputs.version }}
+        run: |
+          set -euo pipefail
+
+          dotnet tool restore
+
+          dotnet vpk '[win]' pack \
+            --skip-updates \
+            --packId DodoSSH.Desktop.Nightly \
+            --packVersion "$VERSION" \
+            --packDir "$RUNNER_TEMP/nightly-win-x64" \
+            --packTitle 'DodoSSH Nightly' \
+            --packAuthors DodoTech \
+            --mainExe DodoSSH.exe \
+            --icon src/DodoSSH.Client.App/Assets/dodossh.ico \
+            --runtime win-x64 \
+            --channel win-nightly \
+            --outputDir "$RUNNER_TEMP/nightly-releases"
+
+          ls -la "$RUNNER_TEMP/nightly-releases"
+
+      # ◆ AND THE LINUX NIGHTLY, WHICH IS A PACKAGE RATHER THAN A FEED.
+      #
+      # Everything above this step is Velopack: an installer, an index file and a client that polls it.
+      # This one is an RPM, and the difference is not a gap in the work. VelopackUpdateChannel refuses
+      # every platform that is not Windows or macOS — deliberately, and its comment says why — so a
+      # Linux build has no self-updater to give a channel to. What it has instead is dnf, which is a
+      # mechanism the user already trusts and already controls, and build/linux/dodossh.spec argues
+      # that at length.
+      #
+      # So there is no rolling *feed* here, only a rolling *file*: the newest nightly RPM sits on the
+      # same release as the Windows nightly and somebody who wants it downloads it. That is a worse
+      # deal than the Windows nightly gets and it is the honest one — the alternative is a package
+      # repository, which is a deployment this project does not have.
+      #
+      # DodoChannel=nightly for the same reason the win-x64 publish above sets it: it is what
+      # ClientPaths reads to suffix the profile directory, so the nightly RPM and a release RPM
+      # installed beside it share no cache, no outbox and no device key. The package name, the tree
+      # under /opt and the desktop entry are the packaging half of that separation and are the
+      # script's `--channel nightly` — see the note there for why it is four things and not one.
+      - name: publish and package the linux nightly
+        id: linuxpack
+        env:
+          VERSION: ${{ steps.version.outputs.version }}
+        run: |
+          set -euo pipefail
+
+          dotnet publish src/DodoSSH.Client.App/DodoSSH.Client.App.csproj \
+            --configuration Release --runtime linux-x64 --self-contained true \
+            -p:RestoreLockedMode=false \
+            -p:DodoChannel=nightly \
+            -p:MinVerVersionOverride="$VERSION" \
+            --output "$RUNNER_TEMP/nightly-linux-x64"
+
+          scripts/ci-linux-rpm.sh \
+            --payload "$RUNNER_TEMP/nightly-linux-x64" \
+            --version "$VERSION" \
+            --channel nightly \
+            --out "$RUNNER_TEMP/nightly-rpm"
+
+      # ◆ PUBLISHING IT, AND WHAT THAT CAPABILITY IS.
+      #
+      # Whoever can write a release here can put a build on every nightly desktop, because Velopack fetches
+      # from this feed and applies what it finds without verifying a signature. That is the same capability
+      # as a signing key reached through a different door, and ADR 0013 rule 3 keeps it off runners.
+      #
+      # It is acceptable here for the reason the android nightly gives, and only for that reason: this is
+      # not that channel. A nightly is a separate application with its own pack id, its own install
+      # directory and its own profile, and it cannot update the build anybody is trusting with their
+      # credentials — the release channel reads a different index and refuses prereleases, so it cannot
+      # even see this one. Anybody installing a nightly is trusting everyone who can write to this
+      # repository, which is a thing to know rather than a thing to discover; the README says so.
+      #
+      # ◆ DELETED AND RECREATED RATHER THAN ADDED TO.
+      #
+      # A rolling channel has to keep exactly one build, and every asset here is version-named, so merging
+      # would grow the release by a hundred and twenty megabytes per push until the forge said no. There is
+      # no atomic form of this in the API, so the shape with the fewest states is to remove both the
+      # release and its tag and make them again. The window where no nightly exists is a few seconds, and
+      # the client's answer to it is the same as to an unreachable forge: the timer swallows it and tries
+      # later; a pressed CHECK NOW says so.
+      - name: publish the nightly release
+        env:
+          FORGE: https://git.dodotech.cloud
+          REPO: DodoTech-Public/DodoSSH
+          TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          VERSION: ${{ steps.version.outputs.version }}
+          TAG: nightly-desktop
+        run: |
+          set -euo pipefail
+
+          if [ -z "${TOKEN:-}" ]; then
+            echo "No token, so the nightly was built and not published." >&2
+            echo "GITHUB_TOKEN is provided by the runner; an empty one means Actions is configured" >&2
+            echo "without it, and the job's contents: write permission is what asks for it." >&2
+            exit 1
+          fi
+
+          api="$FORGE/api/v1/repos/$REPO"
+          auth="Authorization: token $TOKEN"
+
+          # The first "id" with digits after it, for the reason the android job records at length: a
+          # greedy .* walks past the release's own id to the author's, which is -2, and yields nothing.
+          release_id() {
+            grep -oE '"id":[0-9]+' | head -1 | cut -d: -f2
+          }
+
+          existing="$(curl -fsS -H "$auth" "$api/releases/tags/$TAG" 2>/dev/null || true)"
+          if [ -n "$existing" ]; then
+            id="$(printf '%s' "$existing" | release_id)"
+            if [ -n "$id" ]; then
+              echo "Removing the previous $TAG release $id"
+              curl -fsS -X DELETE -H "$auth" "$api/releases/$id" >/dev/null || true
+            else
+              echo "A $TAG release exists and its id could not be read:" >&2
+              printf '%s\n' "$existing" >&2
+              exit 1
+            fi
+          fi
+          curl -fsS -X DELETE -H "$auth" "$api/tags/$TAG" >/dev/null 2>&1 || true
+
+          # vpk rather than curl, unlike the android job above, and the difference is what is being
+          # uploaded. An APK is one file the client is told about by a manifest this repository writes; a
+          # Velopack release is a set of files plus an index whose format is Velopack's own. Writing that
+          # index by hand would be reimplementing the tool that is already here.
+          #
+          # --pre is load-bearing twice over. It keeps this out of the release channel, which refuses
+          # prereleases — and it keeps it out of `releases/latest`, which is what the *phone's* release
+          # channel reads: a desktop nightly published as a stable release would become the newest release
+          # in this repository and every phone on the release channel would start failing its update check
+          # against a release carrying no android manifest.
+          dotnet vpk upload gitea \
+            --skip-updates \
+            --repoUrl "$FORGE/$REPO" \
+            --token "$TOKEN" \
+            --outputDir "$RUNNER_TEMP/nightly-releases" \
+            --channel win-nightly \
+            --tag "$TAG" \
+            --releaseName "Nightly desktop $VERSION" \
+            --targetCommitish "$GITHUB_SHA" \
+            --pre \
+            --publish
+
+          # Asked for back rather than assumed, and the android job's history is why: it once created a
+          # release, uploaded nothing, and reported success for every upload it never made. A nightly
+          # desktop feed that exists and carries no index is a client that checks, finds nothing, and
+          # reports itself up to date forever.
+          published="$(curl -fsS -H "$auth" "$api/releases/tags/$TAG")"
+
+          for name in releases.win-nightly.json DodoSSH.Desktop.Nightly-win-nightly-Setup.exe; do
+            # grep without -q, which has the same exit status and reads to the end of its input. A
+            # reader that stops early under pipefail turns the writer's SIGPIPE into this condition's
+            # answer; see the note in scripts/ci-linux-rpm.sh, where that cost a green build.
+            if ! printf '%s' "$published" | grep -F "\"name\":\"$name\"" >/dev/null; then
+              echo "The release was created but $name is not on it:" >&2
+              printf '%s\n' "$published" >&2
+              exit 1
+            fi
+          done
+
+          # ◆ THE RPM GOES ONTO THE RELEASE vpk JUST MADE, RATHER THAN ONTO ONE OF ITS OWN.
+          #
+          # One rolling release carrying both heads, because they are one nightly: the same commit,
+          # the same version, and one page a person looks at to find out what is on main. A second tag
+          # would double the delete-and-recreate dance this step already does, and would let the two
+          # drift apart on a run where one of them failed.
+          #
+          # It has to happen here, after `vpk upload`, and not in a step before it. vpk deletes and
+          # recreates this release, so an RPM uploaded first would be attached to something that no
+          # longer exists by the time the job ends — silently, because the delete succeeds.
+          #
+          # curl rather than vpk, unlike the assets above, for the reason the android job gives about
+          # its APK: this is one file the release simply carries. There is no index to write and no
+          # format that belongs to a tool.
+          release="$(printf '%s' "$published" | release_id)"
+          if [ -z "$release" ]; then
+            echo "The release was published and its id could not be read back:" >&2
+            printf '%s\n' "$published" >&2
+            exit 1
+          fi
+
+          # ◆ THE PACKAGE FIRST AND THE MANIFEST LAST, WHICH IS THE ORDER THE CLIENT DEPENDS ON.
+          #
+          # LinuxUpdateChannel reads linux-<channel>.json and then fetches what it names, exactly as the
+          # phone reads android-nightly.json. A manifest published before its package is a window in
+          # which every installed client is told to download something that is not there yet — and the
+          # client's answer to a missing asset is to throw, which a pressed CHECK NOW reports as a
+          # failure. Same rule, same reason, and the android job's comment says it first.
+          rpm='${{ steps.linuxpack.outputs.rpm }}'
+          manifest='${{ steps.linuxpack.outputs.manifest }}'
+
+          linuxNames=""
+          for file in "$rpm" "$manifest"; do
+            name="$(basename "$file")"
+            echo "Uploading $name ($(wc -c < "$file") bytes)"
+            curl -fsS -X POST -H "$auth" \
+              -F "attachment=@$file" \
+              "$api/releases/$release/assets?name=$name" >/dev/null
+            linuxNames="$linuxNames $name"
+          done
+
+          # Asked for back, exactly as the Velopack assets above are and for the same history: the
+          # android job once created a release, uploaded nothing, and reported success for every
+          # upload it never made.
+          withRpm="$(curl -fsS -H "$auth" "$api/releases/$release")"
+          for name in $linuxNames; do
+            # grep without -q; see the sibling loops above and scripts/ci-linux-rpm.sh.
+            if ! printf '%s' "$withRpm" | grep -F "\"name\":\"$name\"" >/dev/null; then
+              echo "The release was created but $name is not on it:" >&2
+              printf '%s\n' "$withRpm" >&2
+              exit 1
+            fi
+          done
+
+          echo "Published nightly desktop $VERSION as $TAG, with$linuxNames"
+
   image:
     name: api image
+    # See the build job. A cold layer cache is the slow case here.
+    timeout-minutes: 40
     # Gated on the tests rather than parallel with them, which costs a few minutes of wall
     # clock on every main commit and buys the thing worth having: no image reaches the
     # registry from a commit whose tests were red. An image is not a build artefact anyone
@@ -1067,9 +1598,10 @@ jobs:
         if: always() && github.event_name != 'pull_request'
         run: docker logout registry-docker.dodotech.cloud
 
-# There is no job here that publishes the desktop client, and there is not going to be one. It is built
-# and packaged here — the two steps at the end of the build job — and what is withheld is only the
-# upload.
+# No job here publishes the desktop *release* channel, and there is not going to be one. The desktop
+# nightly job above publishes a different application — its own pack id, its own Velopack channel, its own
+# install directory and profile — and the distance between those two sentences is the whole design. What
+# the build job does for the release channel is prove it still builds and packs; the upload is withheld.
 #
 # ◆ ONE REASON, WHERE THIS ONCE CLAIMED TWO, AND THE SECOND WAS NOT TRUE.
 #

.gitignore

@@ -87,9 +87,16 @@ Releases/
 *.pfx
 *.p12
 *.snk
+# Signing keystores, with one exception that is tracked deliberately. build/nightly.keystore is
+# committed in the open on purpose — see ADR 0014 for why a key everybody holds is what makes the
+# nightly channel safe to run in CI — and a file already tracked is unaffected by an ignore rule
+# anyway. The negation is here so that reading this list does not suggest the tracked one is an
+# oversight. The *release* keystore is a real secret under ADR 0011 rule 1 and must never land here.
+*.keystore
+*.jks
+!build/nightly.keystore
 appsettings.Local.json
 appsettings.*.Local.json
 .env
 .env.*
-!deploy/.env.example
 secrets/

Directory.Build.targets

@@ -0,0 +1,68 @@
+<Project>
+
+  <!--
+    ◆ THE TRIMMER'S VERSION IS PINNED HERE BECAUSE OTHERWISE THE LOCK FILES ARE NOT LOCKED.
+
+    Microsoft.NET.ILLink.Tasks is not referenced by anything in this repository. The SDK adds it
+    on its own to any project that sets IsTrimmable or IsAotCompatible — DodoSSH.Contracts and
+    DodoSSH.Crypto do, and the Android head gets it from trimming being on by default there — and
+    the version it asks for is whatever the running SDK happens to bundle. That version lives in
+    the SDK's own Microsoft.NETCoreSdk.BundledVersions.props, as a KnownILLinkPack item.
+
+    Which makes it a dependency whose version is a property of the toolchain rather than of this
+    repository, and that is the whole problem: packages.lock.json records it as a Direct reference
+    with a requested range, so the lock file silently means "whichever SDK last ran a restore".
+    global.json says rollForward: latestMinor, so CI's setup-dotnet installs the newest 10.x SDK
+    that exists on the day it runs. The moment .NET ships a servicing release, CI's SDK asks for a
+    version the committed lock files do not have, and the locked-mode restore in ci.yml fails with
+    NU1004 before a single file is compiled.
+
+    That is not hypothetical. It closed the whole pipeline: main's run 125 and every open pull
+    request went red together, on
+
+        error NU1004: The package reference Microsoft.NET.ILLink.Tasks version has changed
+        from [10.0.10, ) to [10.0.11, ).
+
+    with nothing in any of those commits touching a package. .NET had shipped SDK 10.0.400, which
+    bundles ILLink 10.0.11 where 10.0.302 bundled 10.0.10, and setup-dotnet installed it the next
+    time anything ran.
+
+    Worse than the outage is the shape of the repair without this pin. Regenerating the lock files
+    holds only until the next servicing release, and it cannot be done from a machine whose newest
+    SDK is older than the runner's: a restore on 10.0.302 writes 10.0.10 straight back and re-breaks
+    CI, so the recorded version becomes a fact about whoever ran restore last rather than about this
+    repository. That is exactly the state locking exists to prevent, and it is not a hypothetical
+    either — every SDK installed on the machine this pin was written on tops out at 10.0.302.
+
+    Pinning it makes the recorded version a decision this repository made, reviewable in a diff
+    like every other version in Directory.Packages.props, and identical on every machine whatever
+    SDK it has. Moving it is then a deliberate edit here plus a regenerated lock file, which is the
+    same ceremony any other dependency bump gets.
+
+    It is an Update on the SDK's item rather than a PackageVersion in Directory.Packages.props, and
+    it has to be: the reference is implicit, so the SDK supplies the version itself and central
+    package management never gets asked. ProcessFrameworkReferences reads @(KnownILLinkPack) when
+    it runs, which is why this lives in Directory.Build.targets — the item does not exist yet while
+    Directory.Build.props is being evaluated.
+
+    Keep this within a patch or two of the runtime the SDK ships. It is the trimming analyzer and
+    the ILLink task, so a small skew is harmless, but a version far behind the framework being
+    analysed is a real way to miss a trim warning.
+  -->
+  <Target Name="PinTheILLinkPackVersion" BeforeTargets="ProcessFrameworkReferences">
+    <!--
+      Inside a target, and not for tidiness. The SDK ships one KnownILLinkPack per target framework
+      and they all share the identity "Microsoft.NET.ILLink.Tasks", so the TargetFramework metadata
+      is the only thing telling net10.0's entry from net8.0's. A condition on %(...) is item
+      batching, which MSBuild permits in a target and rejects during evaluation with MSB4191 — so
+      an ItemGroup at the top of this file cannot express "only the net10.0 one" at all, and the
+      unconditioned Update it would have to become rewrites every framework's entry.
+    -->
+    <ItemGroup>
+      <KnownILLinkPack Update="Microsoft.NET.ILLink.Tasks"
+                       Condition="'%(TargetFramework)' == 'net10.0'"
+                       ILLinkPackVersion="10.0.11" />
+    </ItemGroup>
+  </Target>
+
+</Project>

Directory.Packages.props

@@ -88,8 +88,12 @@
     -->
     <PackageVersion Include="NSec.Cryptography" Version="26.4.0" />
     <PackageVersion Include="libsodium" Version="1.0.22" />
-    <!-- Managed differential oracle for the crypto test suite only. -->
-    <PackageVersion Include="BouncyCastle.Cryptography" Version="2.6.2" />
+    <!--
+      Managed differential oracle for the crypto test suite only. Also what SSH.NET pulls in,
+      so central transitive pinning makes this pin its floor too: 2.7.0 is what SSH.NET
+      2026.0.0 asks for, and pinning below that is a downgrade error rather than a preference.
+    -->
+    <PackageVersion Include="BouncyCastle.Cryptography" Version="2.7.0" />
   </ItemGroup>
   <ItemGroup Label="Desktop client">
     <!--
@@ -98,8 +102,14 @@
       agent forwarding (needs an upstream change; de-scoped from v1) and being handed a
       pre-connected Stream — it performs its own socket connect, which is why the relay and
       ProxyJump both go through a loopback TCP bridge. See docs/adr/.
+
+      2026.0.0 is the fix for GHSA-q939-rpr3-3284, a path traversal in ScpClient.Download's
+      recursive mode that trusts server-supplied names (everything up to and including
+      2025.1.0 is affected). Nothing here uses ScpClient — transfers go through SftpClient —
+      so this was an upgrade on principle rather than an exposure, and the release notes
+      list no breaking changes over 2025.1.0.
     -->
-    <PackageVersion Include="SSH.NET" Version="2025.1.0" />
+    <PackageVersion Include="SSH.NET" Version="2026.0.0" />
     <!--
       The S3 client, for buckets as a remote in the file browser. First-party, Apache-2.0, and
       managed only — no native assets — which is the bar this file sets for anything that gets
@@ -163,6 +173,28 @@
       chosen user to a known-vulnerable build. See docs/adr/0013-desktop-distribution-and-updates.md.
     -->
     <PackageVersion Include="Velopack" Version="1.2.0" />
+
+    <!--
+      Microsoft's own TPM 2.0 stack, TSS.MSR, and the only way LinuxDeviceKeyStore reaches the chip.
+
+      Windows needs nothing here: CNG is in the framework and the Platform Crypto Provider is the TPM
+      behind it, and macOS reaches the Secure Enclave through the Security framework that
+      MacSymbols.cs binds by hand. Linux has neither — the platform's own stack is tpm2-tss, a C
+      library with a large surface — so this is the one head that needs a package to talk to its own
+      hardware.
+
+      Chosen over P/Invoking libtss2-esys for two reasons that both hold at the same time. It is
+      managed all the way down, so there is no native dependency to add to the RPM and nothing to
+      resolve at load time; and it works against a software TPM over a socket exactly as it does
+      against /dev/tpmrm0, which is what makes LinuxDeviceKeyStoreTests able to run the whole seal,
+      unseal and lockout path unattended. Neither of the other two device key stores can be tested at
+      all — see WindowsDeviceKeyStoreTests for the ten minutes that took to find out.
+
+      It carries a runtimes/linux-x64 asset, like SkiaSharp and libsodium, so the RID-specific publish
+      picks it up and the RID-free locked restore does not. See the RuntimeIdentifiers note in
+      DodoSSH.Client.App.csproj for why that distinction is load-bearing here.
+    -->
+    <PackageVersion Include="Microsoft.TSS" Version="2.1.1" />
   </ItemGroup>
   <ItemGroup Label="Versioning">
     <!--
@@ -211,5 +243,20 @@
       mistakes that cause real authorization holes.
     -->
     <PackageVersion Include="WireMock.Net" Version="2.13.0" />
+    <!--
+      A JavaScript interpreter, in managed code, so that terminal.js can be tested as the file that
+      ships rather than as a transcription of it into C#. Test-only and referenced by exactly one
+      project; nothing in src depends on it.
+
+      Chosen over shelling out to node, which is the obvious alternative and needs node present
+      wherever the suite runs — CI installs one already, but a test that only runs when a separate
+      command is remembered is a test that stops being run. Chosen over a headless browser for the
+      same reason several times over.
+
+      What it does not buy: Jint is not Chromium, so this proves the page's own logic and nothing
+      about how WebView2 or Android's WebView behave. That line is drawn in RendererPage's remark and
+      picked up by docs/manual-checks.md 1.10 and 11.12a.
+    -->
+    <PackageVersion Include="Jint" Version="4.16.0" />
   </ItemGroup>
 </Project>

README.md

@@ -69,9 +69,10 @@ src/
   DodoSSH.Client.Terminal  the loopback data plane and credit-based flow control
   DodoSSH.Client.Transfer  the transfer queue, part files and resume, and the local file listing
   DodoSSH.Client.ObjectStore  S3-compatible buckets, behind the same interface as SFTP
-  DodoSSH.Client.Import    reading ~/.ssh/config, and the keys it names when asked
+  DodoSSH.Client.Import    reading ~/.ssh/config and the keys it names when asked; reading a Hetzner
+                           Cloud project over its API
   DodoSSH.Client.Shell     the view models both heads drive, the renderer's files, the palette
-  DodoSSH.Client.App       the desktop head: its views, and its Windows integration
+  DodoSSH.Client.App       the desktop head: its views, and its Windows, macOS and Linux integration
   DodoSSH.Client.Android   the phone head: its views, and its Android integration
 tests/                     one test project per source project
 docs/adr/                  architecture decision records
@@ -80,6 +81,10 @@ docs/platform-flags.md     what differs off Windows, and the gotchas that have c
 docs/manual-checks.md      what no test can reach, and what to look for when checking by hand
 docs/android-port.md       the Android head: what was decided, what is built, what is left
 scripts/                   release-windows.ps1 — builds, packs and publishes the Windows client
+                           release-macos.sh   — the same, signed and notarized, on a Mac
+                           ci-linux-rpm.sh    — builds the Linux .rpm, then interrogates it
+build/macos/               the entitlements and Info.plist template the macOS bundle is built from
+build/linux/               the RPM spec the Linux package is built from
 ```
 
 Everything under `src/DodoSSH.Client.*` except the two heads and `Shell` is deliberately free of Avalonia.
@@ -152,8 +157,129 @@ reinstalling asks for your passphrase rather than starting over. Use **Sign out*
 you want the machine to genuinely forget everything — an uninstall is not a sign-out, and does not withdraw
 this machine's device key from your account.
 
-Cutting a release is `scripts/release-windows.ps1`, run by a person on a Windows machine. Deliberately not a
-CI job; ADR 0013 decision 3 explains why, and it is not only that the runners are Linux.
+## Installing on macOS
+
+A `.pkg` on the same release page, for Apple Silicon. Everything above about where a client may come from,
+about the update check and about uninstalling applies unchanged; what differs is worth three short
+paragraphs.
+
+**It is signed and notarized, so there is no warning to click past.** That is not generosity — macOS refuses
+to open an un-notarized download outright rather than warning about it, so unlike the Windows build there
+was never an unsigned option. If you *do* see "cannot be opened because Apple cannot check it for malicious
+software", the file did not come from the project's release page, and that is worth taking literally.
+
+**Apple Silicon only for now.** An Intel package is a small amount of work and no one here has an Intel Mac
+to check it on, and this project does not ship desktop builds nobody has run — see
+[docs/manual-checks.md](docs/manual-checks.md). Under Rosetta the arm64 build will not run; there is no
+graceful version of that, and the honest answer is that the platform is not covered yet.
+
+**Touch ID can stand in for your passphrase**, on a Mac with a Secure Enclave. The key that unwraps your
+device key is generated inside the enclave and never leaves it, and the enclave — not DodoSSH — is what
+requires your fingerprint or login password before it will use it. Cancel the prompt and you get the
+passphrase screen, always. The application lives at `/Applications/DodoSSH.Desktop.app` and your vault cache
+at `~/Library/Application Support/DodoSSH`, which are deliberately two different places so that removing the
+first never touches the second.
+
+## Installing on Linux
+
+An `.rpm` for x86_64, and the one platform where you get it from `main` rather than from a release — see
+[the nightly build](#the-nightly-desktop-build) below, which is where the Linux package is published today.
+
+```bash
+sudo dnf install ./dodossh-nightly-*.x86_64.rpm
+```
+
+Everything above about where a client may come from applies unchanged. Four things differ, and the first
+two are the ones to read before installing.
+
+**It checks for updates and it installs nothing.** The Windows and macOS builds carry Velopack and replace
+their own files; this one never does. It checks the project's forge every six hours, and when there is a
+newer build it downloads it, checks its digest, and hands the file to **PackageKit** — which raises the
+usual "authentication is required to install software" prompt. You may say no, and nothing is replaced until
+you say yes. Nothing restarts either: replacing files under a running process is ordinary here, so the build
+you have keeps running until you close it.
+
+That split is the point rather than a compromise. Installing software is your package manager's job and it
+already answers to you; noticing that a build exists is the part `dnf` cannot do for a package that is not
+in a repository. What DodoSSH must never become is a second thing that can put software on your machine on
+somebody else's say-so, which is why the address it checks is a constant in the binary and can never be a
+DodoSSH server — [ADR 0013](docs/adr/0013-desktop-distribution-and-updates.md) rules 4 and 11.
+
+The gap that remains: there is **no package repository**, so `dnf upgrade` alone will never bring you a
+newer DodoSSH. If you turn the check off on PREFERENCES → UPDATES, nothing else will tell you.
+
+**Your PIN can stand in for your passphrase, if this machine has a TPM you can reach.** The device key is
+sealed into the TPM and released only against a PIN that **the TPM itself checks** — it counts wrong answers
+and stops responding after a few, which is what makes four digits enough and what makes a stolen disk
+useless. Two honest caveats, both in
+[ADR 0007](docs/adr/0007-device-key-protection.md) decision 4:
+
+- **It is weaker than Windows Hello or Touch ID**, and the difference does not show. Those draw their prompt
+  from inside the operating system; this one is a window DodoSSH draws, so something already running as you
+  could imitate it. It could imitate the passphrase screen too, which is worth far more — but the guarantee
+  is not the same and this page will not pretend it is.
+- **The package grants your session access to the TPM, and you should know what that means.** Fedora leaves
+  `/dev/tpmrm0` readable only by the `tss` group, so this would otherwise need an administrator and a fresh
+  login. Instead the package installs a udev rule tagging the device `uaccess` — the same mechanism your
+  webcam and your FIDO key already use, which grants an ACL to **whoever is logged in at the machine** and
+  takes it away when they log out. Not over SSH, not to a user switched away on another VT, not to daemons.
+
+  The cost is not secrecy — TPM objects are protected by their own PINs — it is that anyone at the keyboard
+  can make failed attempts, and a TPM answers enough of those by locking out *everyone*. **If this machine
+  unlocks its disk with the TPM at boot**, that means somebody sitting at it could stop the next boot from
+  unlocking until the lockout recovers. If you would rather not have that, one command puts it back:
+
+  ```bash
+  sudo ln -s /dev/null /etc/udev/rules.d/71-dodossh-tpm.rules
+  ```
+
+  DodoSSH then behaves as though the machine had no TPM: no PIN offered, passphrase every time.
+
+See [docs/unlocking-without-the-passphrase.md](docs/unlocking-without-the-passphrase.md).
+
+**It needs WebKitGTK, because the terminal is a WebView.** Specifically `webkit2gtk4.1`, which `dnf` pulls
+in for you. Fedora 44 is the distribution the head has actually been run on. RHEL 9 and its rebuilds ship
+the older 4.0 API line under a different library name and the package deliberately does not claim them —
+`dnf` will refuse with *nothing provides libwebkit2gtk-4.1.so.0*, which is a better answer than installing
+and drawing a blank terminal. This project does not ship desktop builds nobody has run, on any platform.
+
+**Where things go.** The application at `/opt/DodoSSH`, launched as `dodossh` or from your menu; your vault
+cache at `$XDG_DATA_HOME/dodossh`, usually `~/.local/share/dodossh`. Two different places on purpose:
+`dnf remove` takes the first and never touches the second, so reinstalling asks for your passphrase rather
+than starting over. Use **Sign out** inside the application to make the machine genuinely forget.
+
+Cutting a release is `scripts/release-windows.ps1`, run by a person on a Windows machine, and
+`scripts/release-macos.sh` on a Mac. Deliberately not a CI job; ADR 0013 decision 3 explains why, and it is
+not only that the runners are Linux.
+
+### The nightly desktop build
+
+There is a second Windows build, published by CI from `main` on every push, on the `nightly-desktop`
+release. It installs **beside** the release build rather than over it — its own entry called *DodoSSH
+Nightly*, its own install directory, and its own profile at `%LOCALAPPDATA%\DodoSSH.Nightly` — so trying it
+costs nothing you are relying on. It signs in separately and appears to your account as a new device,
+because it genuinely is a second installation.
+
+**The Linux `.rpm` is on the same release**, built from the same commit and versioned the same way, as
+`dodossh-nightly`. It installs beside a release package for the same reasons and by the same four
+separations — a different package name, a tree at `/opt/DodoSSH.Nightly`, a `dodossh-nightly` launcher with
+its own menu entry, and a profile at `~/.local/share/dodossh.nightly`. Two RPMs owning different paths
+cannot replace one another, so `dnf` will keep both rather than treating one as an upgrade of the other.
+
+It is the only Linux package this project publishes at all today: there is no `v*`-tagged Linux release
+yet, because no person has walked a Linux build through
+[docs/manual-checks.md](docs/manual-checks.md). What CI proves on every push to `main` is that the release
+package still *builds* — the same arrangement Windows and macOS have had since the beginning, where the
+package is built on every commit and uploaded nowhere.
+
+**Installing it means trusting everyone who can write a release on this repository, including CI.** The
+updater applies what the feed serves without verifying a signature, so anyone who can change a workflow
+file here can put a build on every nightly machine. That is an acceptable trade for a build you are trying
+and not one for a build holding your infrastructure credentials, which is the entire reason the two
+channels exist and cannot see each other. The release build refuses prereleases and reads a different feed,
+so nothing published here can ever reach it. Same arrangement, and same reasoning, as the phone's nightly —
+[ADR 0014](docs/adr/0014-android-updates.md), and [ADR 0013](docs/adr/0013-desktop-distribution-and-updates.md)
+decision 9.
 
 ## Running it
 
@@ -407,8 +533,28 @@ The two vaults are encrypted under different keys, so a move is a re-seal into o
 other; the host gets a new id, and **its group and its tags stay behind**, because both are items of the
 vault it is leaving. A picker inside the form would do all of that as a side effect of correcting a port.
 What a move cannot do is reach a machine that has already synced the host, which is the same limit
-everything else about revocation has. Keys, passwords and buckets take theirs from a standing "new items go
-to" picker on the Keychain screen and cannot be moved yet.
+everything else about revocation has. Buckets take theirs from a standing "new items go to" picker on the
+Keychain screen and cannot be moved yet.
+
+**The move asks whether the key comes too**, because that is the half a host's move could not settle on its
+own. A binding resolves across every vault you can read, so the moved host goes on working for *you* either
+way — but the people you have just shared it with hold one vault's key, and a host whose key stayed in your
+personal vault is one they cannot connect with. The tick box beside the picker is unticked, and stays that
+way on purpose: moving a key into a team's vault hands it to everybody who holds that key, and a disclosure
+is chosen rather than defaulted into. Under it is the count of everything else that authenticates with that
+key, which is what makes the answer decidable — a key twenty machines use is a different decision from one
+nothing else touches. Left unticked, the sentence afterwards names the key that is now outside the
+destination. A key the host only *inherits* from its group counts too, and is written onto the host on the
+way across: the group stays behind, so a host that arrived naming nothing would authenticate with nothing.
+
+**A key or a password can also be moved on its own** — MOVE beside EDIT and DELETE on the Keychain screen,
+on both heads, for keys and passwords only. It is the same re-seal and tombstone, and it takes a new id in
+the destination, so **everything that named it is re-aimed at where it went**: every host bound to it and
+every group lending it as a default, across every vault you can write to. Without that the move would be a
+deletion with extra steps, since a host bound to something its vault no longer holds refuses to connect
+rather than falling back to a typed password. Anything that cannot be rewritten here — an item from a newer
+client, or one in a vault you can only read — is left naming the old item and is counted in the sentence
+afterwards. The panel says what points at the key before you press it, not after.
 
 **A group can be moved too, and it takes its contents with it** — "Move to another vault…" on the group
 card's right-click menu, beside Open, Edit and Delete, which is the whole of what can be done to a group on
@@ -727,6 +873,13 @@ keychain plus a terminal — and the spike that gates all of it.
   [Moving files](#moving-files) for the two things about it worth knowing before you use it, both of which
   are consequences rather than choices.
 
+  *Relay done:* a host with "Connect through the server relay" ticked is reached through the deployment,
+  for a machine this one cannot dial. The server issues a single-use, thirty-second ticket for that host
+  alone, resolves its address and checks it against a deny list with no configuration escape — loopback,
+  link-local, metadata endpoints — and then pipes raw TCP; the SSH session still ends on your machine, so
+  the relay sees ciphertext. It is off by default (`Relay:Enabled`), and it only helps when the server can
+  reach what you cannot. See [ADR 0004](docs/adr/0004-relay-authorization.md).
+
   *Organising done:* hosts can be filed into groups, and commands can be saved as snippets. Both are ordinary
   synced items — encrypted, merged and pushed like every other — and both are invisible until used: a
   keychain with no groups draws the flat host list it always did. Two things about them are deliberate.
@@ -849,8 +1002,39 @@ keychain plus a terminal — and the spike that gates all of it.
   [ADR 0013](docs/adr/0013-desktop-distribution-and-updates.md), and
   [Installing on Windows](#installing-on-windows) for what a user sees.
 
-  Still to do here: signing (the first release is unsigned, and the trigger for buying a certificate is the
-  first release aimed at strangers), and macOS and Linux packaging.
+  **The macOS half is built on the same machinery**, and signed from the start because Gatekeeper leaves no
+  choice: `scripts/release-macos.sh` publishes, signs every native library, notarizes with Apple and staples
+  the ticket before it will hand anything over, and refuses to upload until a person has installed it. The
+  device key is held in the Secure Enclave behind Touch ID. CI publishes `osx-arm64` and builds the `.app`
+  on every main build to prove it still packages, and uploads nothing. See
+  [ADR 0013](docs/adr/0013-desktop-distribution-and-updates.md) decision 10 and
+  [Installing on macOS](#installing-on-macos).
+
+  **The Linux half checks, and the package manager installs.** An `.rpm` for x86_64, built and checked on
+  every push to `main` and published to the nightly release, installing to `/opt` with a menu entry and a
+  `dodossh` launcher. It carries no Velopack and replaces nothing itself: it notices a newer build on the
+  project's forge, downloads it, verifies its digest and hands the file to PackageKit, which asks before it
+  installs. Splitting it that way keeps the part that can put software on your machine where it already
+  was — with your package manager — while fixing the half `dnf` cannot do for a package that is in no
+  repository. The device key is sealed into the TPM behind a PIN **the TPM checks and counts failures
+  against**, which is what makes four digits enough. Reaching the TPM needs no administrator and no group:
+  the package installs a udev rule tagging `/dev/tpmrm0` with `uaccess`, so the ACL follows whoever is
+  logged in at the machine and goes away with them — the same mechanism a webcam uses. What that costs, and
+  the one symlink that undoes it, is in the rule's own comments. See
+  [ADR 0007](docs/adr/0007-device-key-protection.md) decision 4,
+  [ADR 0013](docs/adr/0013-desktop-distribution-and-updates.md) decision 11 and
+  [Installing on Linux](#installing-on-linux).
+
+  Still to do here: Windows signing (the first Windows release is unsigned, and the trigger for buying a
+  certificate is the first release aimed at strangers), macOS on Intel, a `v*`-tagged Linux release and a
+  signed package repository to serve it from, and Phases 18 and 19 of the manual checks — neither the macOS
+  nor the Linux build has been walked through by a person. There is no macOS runner in CI at all. The Linux
+  package is verified as far as what it *says* about itself — CI builds it and interrogates its
+  dependencies, its file list and its modes, and installs nothing — and the two pieces of Linux behaviour
+  that are code rather than packaging do have tests: the update channel's version comparison and digest
+  check run in CI, and the TPM store's seal, unseal, wrong-PIN and lockout paths run against a software TPM
+  anybody can start with one `docker run`. What no runner covers is `/dev/tpmrm0`, PackageKit and polkit,
+  which is what Phase 19 is for.
 - **M5 — multi-provider OIDC**, identity key rotation, per-item content keys.
 
 ## Licence

build/linux/71-dodossh-tpm.rules

@@ -0,0 +1,54 @@
+# Lets the person sitting at this machine use its TPM, without joining a group.
+#
+# ◆ WHAT THIS CHANGES, AND WHAT IT DELIBERATELY DOES NOT.
+#
+# Fedora's 60-tpm-udev.rules leaves /dev/tpmrm0 as root:tss mode 0660, so using it means being in
+# the tss group — which means an administrator, a usermod and a fresh login before DodoSSH will
+# offer to seal a device key. This adds the uaccess tag, and nothing else: the mode and the group
+# are untouched, so everything that worked before still works and an ACL is layered on top.
+#
+# uaccess is systemd's own mechanism for exactly this question and is what /dev/dri, webcams,
+# scanners and FIDO keys already use. logind grants a POSIX ACL to the user of the *active local
+# session* and takes it away again when that session ends. So:
+#
+#   - somebody logged in at the physical machine gets access;
+#   - somebody logged in over SSH does not;
+#   - a user switched away to another VT does not;
+#   - a system daemon running as its own user does not.
+#
+# That is a much narrower grant than adding a person to tss, which is permanent, applies over SSH,
+# and applies whether or not they are the one at the keyboard.
+#
+# ◆ tpmrm AND NOT tpm, WHICH IS THE WHOLE REASON THIS IS DEFENSIBLE.
+#
+# /dev/tpm0 is the raw device and access to it is exclusive — two users cannot share it, and a
+# process holding it can leave the chip in a state that breaks everything else. /dev/tpmrm0 is the
+# kernel's resource manager: it virtualises handles, swaps contexts and isolates callers from each
+# other, and it exists *because* the TPM is meant to be usable by more than one unprivileged
+# program. Granting a seat's user the manager is the use it was built for. Granting the raw device
+# would not be, and this rule never names it.
+#
+# ◆ THE COST, WHICH IS REAL AND IS NOT ABOUT SECRECY.
+#
+# TPM keys are protected by their own auth values, so this grants nobody the ability to read
+# anybody else's sealed data. What it does grant is the ability to make failed authorisation
+# attempts, and a TPM answers those by counting them and eventually refusing everyone — the same
+# dictionary-attack lockout DodoSSH relies on to make a four-digit PIN safe.
+#
+# On a machine whose disk is unlocked by the TPM at boot (systemd-cryptenroll, Clevis), a local user
+# who trips that lockout can stop the next boot from unlocking until it recovers. That is a denial
+# of service against the machine's owner by somebody already sitting at the machine, and it is worth
+# knowing rather than discovering. An administrator who does not accept it can mask this file:
+#
+#     sudo ln -s /dev/null /etc/udev/rules.d/71-dodossh-tpm.rules
+#
+# which is the standard override and survives upgrades. DodoSSH then behaves exactly as it did
+# before this file existed: no device key is offered, and unlocking asks for the passphrase.
+#
+# Numbered 71 because 73-seat-late.rules is what turns the tag into an ACL, and 60-tpm-udev.rules
+# is what creates the device node's ownership in the first place. Anything outside that gap either
+# never runs or runs too late to matter.
+
+ACTION=="remove", GOTO="dodossh_tpm_end"
+SUBSYSTEM=="tpmrm", KERNEL=="tpmrm[0-9]*", TAG+="uaccess"
+LABEL="dodossh_tpm_end"

build/linux/dodossh.spec

@@ -0,0 +1,342 @@
+# ==================================================================================================
+#  The Linux desktop head, as an RPM.
+#
+#  Built by scripts/ci-linux-rpm.sh, which is what supplies every dodo_* macro below and what checks
+#  the result. Nothing here is invoked by hand; `rpmbuild -bb` on this file with no --define set will
+#  fail on the Version line, which is the intended way round.
+#
+#  ◆ WHY THIS IS AN RPM AND NOT A VELOPACK RELEASE, WHICH IS THE FIRST QUESTION.
+#
+#  Both desktop platforms that ship today are Velopack: the client polls the project's forge, fetches
+#  a newer build and swaps itself over. ADR 0013 rule 4 is what makes that safe to hold — the feed
+#  address is a const, so no operator can answer "is there a newer version" on the user's behalf.
+#
+#  This package replaces nothing and installs nothing, and that is deliberate rather than unfinished.
+#  On Linux the package manager is already an updater the user trusts, audits and controls, so putting
+#  a second one inside the application — answering to this project instead — would be building the
+#  capability ADR 0013 rule 4 spends a page making unsettable, on the one platform where nobody had to.
+#
+#  What the client does do is the half a package manager cannot: LinuxUpdateChannel notices that a
+#  newer build exists on the project's forge, fetches it, checks its digest, and hands the file to
+#  PackageKit — which raises polkit and asks. The user may say no, and nothing is replaced until they
+#  say yes. See ADR 0013 rule 11.
+#
+#  The consequence to know: discovery is worse here than on the other two desktops, because there is
+#  no package repository to subscribe to and the in-app check is what stands in for one.
+# ==================================================================================================
+
+# ◆ THE THREE THINGS THAT MUST BE TURNED OFF, AND WHAT EACH ONE WOULD DO.
+#
+# __os_install_post is the chain of brp-* scripts rpm runs over the buildroot: strip, strip-static,
+# compress, elfperms. Every one of them is wrong here. The payload is a published .NET tree whose
+# managed PDBs ADR 0013 rule 6 keeps on purpose — a stack with file names and line numbers is this
+# project's entire diagnostic channel — and brp-strip would take the native symbols with them for a
+# saving nobody asked for. It also runs the host's binutils over glibc binaries on a build host that
+# is musl, which is a class of breakage with no good failure message.
+#
+# debug_package would then try to build a -debuginfo subpackage out of what strip extracted, from
+# third-party binaries nobody here will ever step through.
+#
+# _build_id_links is the one that surprises people, because turning the other two off does not turn
+# it off: rpm still writes /usr/lib/.build-id/<xx>/<hash> symlinks for every ELF file in the payload.
+# For a bundled tree that means dozens of links in a system directory this package has no business
+# owning, and two files with the same build id — which a .NET publish does produce — make it a build
+# *warning* about duplicates rather than an error, so it goes unnoticed.
+%global __os_install_post %{nil}
+%global debug_package     %{nil}
+%global _build_id_links   none
+
+# ◆ AUTO-REQUIRES STAYS ON. AUTO-PROVIDES DOES NOT. THE ASYMMETRY IS THE WHOLE POINT.
+#
+# rpm's elfdeps reads DT_NEEDED and the symbol-version table out of every ELF file in the buildroot,
+# and the requires half of that is worth much more than a hand-written list: it produces
+# `libc.so.6(GLIBC_2.38)(64bit)` rather than `glibc`, so the package refuses to install on a host
+# whose glibc is too old instead of installing and dying at the first launch. A hand list cannot
+# express that and would go stale the first time a dependency changed underneath it.
+#
+# It reads ELF directly rather than by loading anything, so it does this correctly for x86_64 glibc
+# binaries on a musl build host — which is what lets this package be built on the same Alpine runner
+# as everything else in ci.yml.
+#
+# The provides half is the opposite. Everything under /opt here is a *bundled* copy: libSkiaSharp,
+# libcoreclr, libe_sqlite3 and the rest belong to this application and to nothing else. Left on,
+# this package would advertise `libcoreclr.so()(64bit)` to the whole system, and some other package's
+# dependency could then be satisfied by a private copy inside a directory `dnf remove dodossh` takes
+# away. Excluding the tree by path is the standard shape for that and costs nothing, because nothing
+# outside this package is meant to link against anything inside it.
+#
+# ◆ WHICH IS WHY THE REQUIRES SIDE NEEDS AN EXCLUSION TOO, AND WHY IT IS COMPUTED.
+#
+# The two halves are not independent. A bundled library with a DT_NEEDED on another bundled library
+# produces a requirement for a soname this package carries and — after the exclusion above — no
+# longer advertises, which no host on earth can satisfy. It is not hypothetical and it is not rare:
+# .NET's own libmscordbi.so links libmscordaccore.so, so the very first realistic payload this was
+# tried against produced a package `dnf install` refuses outright.
+#
+# dodo_bundled is the alternation of every lib*.so* filename in the payload, worked out by
+# scripts/ci-linux-rpm.sh from the tree it was handed rather than listed here. A list in this file
+# would be a list of Microsoft's and SkiaSharp's internal file names, kept in step by hand, and its
+# failure mode is the uninstallable package above.
+#
+# liblttng-ust is excluded on its own terms and not because it is bundled — it is not. It is what
+# libcoreclrtraceptprovider.so links, and that file exists so that CoreCLR can emit LTTng trace
+# events if a host has the tracing stack. Nothing here asks it to. The file is kept and the
+# requirement is dropped, which is the arrangement that works both ways round: where liblttng-ust is
+# present the dlopen succeeds and tracing is available, and where it is absent CoreCLR carries on, as
+# it does on every machine that runs a .NET application without it. Requiring it instead would put
+# the package out of reach on RHEL, where lttng-ust is in EPEL rather than in the base repositories —
+# an install failure on a supported distribution, in exchange for a tracing integration this
+# application does not use.
+#
+# scripts/ci-linux-rpm.sh checks the finished package against its own file list afterwards, because
+# an exclusion that silently stopped matching would put the uninstallable package back with no
+# warning anywhere.
+%global __provides_exclude_from ^/opt/.*$
+%global __requires_exclude      ^(liblttng-ust\.so|%{dodo_bundled})
+
+Name:           %{dodo_name}
+Version:        %{dodo_version}
+Release:        %{dodo_release}
+Summary:        SSH client with an end-to-end encrypted credential vault
+License:        MIT
+URL:            https://git.dodotech.cloud/DodoTech-Public/DodoSSH
+BuildArch:      x86_64
+
+# ◆ THE DEPENDENCIES elfdeps CANNOT SEE, WHICH ARE THE ONES THE PRODUCT IS MADE OF.
+#
+# Everything below is reached through dlopen, by a name composed at runtime, so it appears in no
+# DT_NEEDED entry anywhere in the payload. Left to the automatic scan the package would install
+# cleanly on a host with none of it and fail at the first window.
+#
+# GTK and WebKitGTK are the terminal. Avalonia.Controls.WebView P/Invokes a bare `libwebkit2gtk` and
+# a bare `libgtk` and resolves them itself, and MainWindow.OnTerminalEnvironmentRequested documents
+# the backend it was measured against: WebKitGTK 2.52.5 on Fedora 44, which is the 4.1 API line — the
+# assembly's one fully-spelled soname, libjavascriptcoregtk-4.1.so.0, is that line's other half.
+#
+# ◆ THE 4.0 LINE IS DELIBERATELY NOT CLAIMED, AND THE COST OF THAT IS KNOWN.
+#
+# The control's own diagnostic says "webkit2gtk 4.0+", so a dependency permitting either line is
+# tempting and was written first. RHEL 9 and its rebuilds are what it would buy: they ship
+# webkit2gtk3, which is the 4.0 API under the soname libwebkit2gtk-4.0.so.37 — not .so.0, which is
+# what makes this a thing to look up rather than guess. Requiring only 4.1 means `dnf install` on
+# AlmaLinux 9 refuses with "nothing provides libwebkit2gtk-4.1.so.0", by design.
+#
+# It is refused because nobody has run it. This project does not ship an Intel macOS package for the
+# same reason and says so in ADR 0013 decision 10 — the engineering is small and the verification is
+# the scarce thing — and an install that succeeds and then draws a blank terminal is a worse answer
+# than one that refuses in a sentence naming the missing library. The day somebody runs the head on
+# a 4.0 host, this becomes a rich `or` and a line in the README.
+#
+# libX11 is Avalonia's own backend, P/Invoked the same way. On Wayland this is satisfied by XWayland,
+# which every desktop shipping this package has.
+#
+# libicu is the one that looks removable and is not. DodoSSH.Client.App sets InvariantGlobalization
+# to false — deliberately, and the csproj says why: this process prints timestamps and host names to
+# a person who expects their own locale's conventions. A .NET process configured that way and given
+# no ICU does not degrade to invariant, it refuses to start. The bare package name is the Fedora and
+# RHEL spelling; it is the one name here that is a distribution's rather than a soname's, because ICU
+# bumps its soname on every release and pinning one would make this package expire.
+Requires:       libwebkit2gtk-4.1.so.0()(64bit)
+Requires:       libgtk-3.so.0()(64bit)
+Requires:       libgdk-3.so.0()(64bit)
+Requires:       libgio-2.0.so.0()(64bit)
+Requires:       libglib-2.0.so.0()(64bit)
+Requires:       libgobject-2.0.so.0()(64bit)
+Requires:       libX11.so.6()(64bit)
+Requires:       libicu
+
+# The size directories under /usr/share/icons/hicolor belong to hicolor-icon-theme and not to this
+# package, which is the right way round — every application drops icons into them. Depending on the
+# owner is how those directories come to exist at all on a minimal host, and how they get removed
+# again when the last thing using them goes.
+Requires:       hicolor-icon-theme
+
+# ◆ RECOMMENDED RATHER THAN REQUIRED, WHICH IS A JUDGEMENT ABOUT WHAT THE PACKAGE IS FOR.
+#
+# PackageKit is what LinuxUpdateChannel hands a downloaded update to, so without it the update path
+# stops at the last step. It is not required, for two reasons that both point the same way: dnf
+# installs weak dependencies by default, so the ordinary Fedora install gets it anyway; and a machine
+# deliberately built without it — a minimal or an immutable one — should still be able to install an
+# SSH client. The failure is not silent, which is what makes this safe: the channel names the missing
+# service and the file it fetched, so the answer is one command the user was going to run anyway.
+#
+# Nothing is required for the TPM. LinuxDeviceKeyStore talks to /dev/tpmrm0 through a managed stack,
+# so there is no library to depend on, and a machine with no TPM — or one whose device node this user
+# may not open, which is the ordinary Fedora case — simply keeps asking for the passphrase.
+Recommends:     PackageKit
+
+%description
+DodoSSH is an SSH and SFTP client for teams whose credentials are shared and whose infrastructure is
+not. Hosts, keys and passphrases live in a vault that is encrypted on this machine and stays
+encrypted on the server, so the deployment holding a team's secrets can never read them.
+
+This package carries its own .NET runtime. It checks the project's own forge for newer builds and can
+download one, but it installs nothing itself: an update is handed to PackageKit, which asks before it
+replaces anything. The address it checks is fixed and is never a DodoSSH server.
+
+%install
+rm -rf %{buildroot}
+
+# ◆ /opt, AND NOT /usr/lib64/dodossh, WHICH IS THE MORE CONVENTIONAL PLACE.
+#
+# The Filesystem Hierarchy Standard reserves /opt for exactly this — an add-on package that arrives
+# as one self-contained tree — and that is what a published .NET application is: a runtime, thirty
+# managed assemblies and half a dozen native libraries that only make sense together and are only
+# ever loaded by each other. Scattering them into the system library directory would mix a bundled
+# libSkiaSharp in with the system's shared objects, which is the arrangement the provides exclusion
+# above exists to keep this package out of.
+#
+# The directory name carries the channel, so a release and a nightly are two trees rather than one
+# that overwrites the other. That is the same separation ADR 0013 rule 9 builds on Windows out of
+# pack ids, arrived at here for free: two RPMs with different names owning different paths cannot
+# replace one another, and dnf will say so rather than doing it. The profile directory is separated
+# by the build itself — DodoChannel=nightly is what ClientPaths reads to suffix it — so the two
+# installs share no state at all.
+mkdir -p %{buildroot}/opt/%{dodo_dir}
+cp -a %{dodo_payload}/. %{buildroot}/opt/%{dodo_dir}/
+
+# ◆ AND THE MODES ARE SET HERE RATHER THAN INHERITED, WHICH IS NOT TIDINESS.
+#
+# `cp -a` preserves what the publish directory happened to have, and what it happened to have is a
+# property of the filesystem it was written on rather than of anything decided here. rpmlint caught
+# the consequence on the first tree this was tried against: every file in the package was mode 777,
+# because the tree had come off a mount that reports everything that way.
+#
+# For this application that is not a lint warning, it is a hole. A world-writable /opt/DodoSSH lets
+# any local account replace the SSH client belonging to every other account on the machine — the
+# binary that holds a team's infrastructure credentials in plaintext while it runs, which is the
+# thing ADR 0001 spends its whole trust model on. Malware running as nobody-in-particular would not
+# need to attack the crypto; it would edit the client.
+#
+# Stated absolutely and in this order, so the answer does not depend on the publish at all:
+# directories and the things that are executed or loaded get 0755, and everything else — assemblies,
+# PDBs, the runtime config, the terminal's HTML and JavaScript — gets 0644. Owned by root, which is
+# rpm's default for a package that names no %attr user, and is what makes the modes mean anything.
+find %{buildroot}/opt/%{dodo_dir} -type d -exec chmod 0755 {} +
+find %{buildroot}/opt/%{dodo_dir} -type f -exec chmod 0644 {} +
+find %{buildroot}/opt/%{dodo_dir} -type f -name '*.so*' -exec chmod 0755 {} +
+
+# The two files a .NET publish means to be executed. createdump is the runtime's crash-dump helper
+# and is only present on some publishes, so it is conditional; the apphost never is, and the check
+# above this section has already refused a publish without one.
+chmod 0755 %{buildroot}/opt/%{dodo_dir}/DodoSSH
+[ -f %{buildroot}/opt/%{dodo_dir}/createdump ] && chmod 0755 %{buildroot}/opt/%{dodo_dir}/createdump
+true
+
+# The apphost is the thing to launch and every other file in the tree is loaded relative to it, so
+# what goes on PATH is a symlink rather than a copy or a wrapper script. .NET resolves its own
+# directory through /proc/self/exe, which follows the link, so a symlinked apphost finds its runtime
+# with no help — a wrapper exporting DOTNET_ROOT or cd-ing first would be solving a problem that
+# does not exist here, and would be one more file to keep in step with the layout.
+#
+# Relative — ../../opt from /usr/bin — rather than the absolute /opt path it resolves to. Both work
+# on a running system and only one works inside a chroot or an image being assembled somewhere other
+# than /, which is where absolute symlinks out of a package are checked and warned about.
+mkdir -p %{buildroot}%{_bindir}
+ln -s ../../opt/%{dodo_dir}/DodoSSH %{buildroot}%{_bindir}/%{dodo_bin}
+
+# ◆ THE DESKTOP ENTRY IS WRITTEN HERE RATHER THAN SHIPPED AS A SOURCE FILE.
+#
+# Three of its seven meaningful lines differ between the release and the nightly, and the two files
+# that would result are the kind of pair that drifts: somebody adds a MIME type to one of them and
+# nobody notices for a release. Composed from the same macros the rest of the layout uses, they
+# cannot.
+#
+# StartupWMClass is the line that is easy to leave out and visibly wrong without: Avalonia sets
+# WM_CLASS from the assembly name, which is DodoSSH for both channels, so without this a launched
+# window is a second, unlabelled icon in the dock beside the one that launched it. It is the assembly
+# name deliberately and not the package's binary name — the window does not know which package
+# started it.
+mkdir -p %{buildroot}%{_datadir}/applications
+cat > %{buildroot}%{_datadir}/applications/%{dodo_bin}.desktop <<DESKTOP
+[Desktop Entry]
+Type=Application
+Version=1.0
+Name=%{dodo_title}
+GenericName=SSH Client
+Comment=SSH and SFTP with an end-to-end encrypted credential vault
+Exec=%{dodo_bin}
+Icon=%{dodo_bin}
+Terminal=false
+StartupNotify=true
+StartupWMClass=DodoSSH
+Categories=Network;RemoteAccess;System;
+Keywords=SSH;SFTP;Terminal;Remote;Shell;
+DESKTOP
+
+# One file per size into the hicolor theme, named for the package so the two channels show different
+# entries rather than fighting over one path. The sizes are the ones the icon script draws and its
+# comment argues for; adding one there means adding it here.
+for size in 48 64 128 256; do
+    mkdir -p %{buildroot}%{_datadir}/icons/hicolor/${size}x${size}/apps
+    install -m 0644 %{dodo_icons}/dodossh-${size}.png \
+        %{buildroot}%{_datadir}/icons/hicolor/${size}x${size}/apps/%{dodo_bin}.png
+done
+
+# ◆ THE UDEV RULE, WHICH IS THE ONE FILE THIS PACKAGE PUTS OUTSIDE ITS OWN TREE.
+#
+# It tags /dev/tpmrm0 with uaccess so the person at the machine can use the TPM without being put in
+# the tss group by an administrator. The rule itself carries the whole argument, including what it
+# costs and the one-line way to mask it.
+#
+# Installed for both channels, and that is not a mistake to be tidied up. The two packages own
+# different paths for the same reason ADR 0013 rule 9 gives, but this file is about the machine's
+# TPM rather than about either build — a nightly and a release on one machine want exactly the same
+# access, and giving each its own copy would mean removing one package silently revoked the other's.
+# rpm refuses to let two packages own one path with different content, which is what makes this
+# safe: the file is byte-identical from both, so it is shared rather than duplicated.
+#
+# /usr/lib/udev/rules.d hardcoded rather than %{_udevrulesdir}, which comes from systemd-rpm-macros
+# and is not present on the Alpine host CI builds this on. Hardcoding a path that has been stable
+# for a decade beats a macro that expands to nothing and silently installs into the buildroot's root.
+mkdir -p %{buildroot}/usr/lib/udev/rules.d
+install -m 0644 %{dodo_udev} %{buildroot}/usr/lib/udev/rules.d/71-dodossh-tpm.rules
+
+# Installed rather than listed straight out of the source tree. `%license` is a %doc-style marker and
+# resolves its argument against the build directory, which this spec does not have — there is no
+# %prep and no %build, because the payload arrives already published. Putting the file where it
+# belongs first and marking the installed path second is the form that works with neither.
+mkdir -p %{buildroot}%{_licensedir}/%{name}
+install -m 0644 %{dodo_licence} %{buildroot}%{_licensedir}/%{name}/LICENSE
+
+# Modern rpm on Fedora fires file triggers that do both of these, so on the distribution this was
+# measured against they are redundant. They are here for the ones that do not, and every line ends in
+# `|| :` so that a host without the tool installed gets a package that installs rather than a
+# scriptlet failure at the end of a 130 MB transaction.
+%post
+/usr/bin/update-desktop-database %{_datadir}/applications &>/dev/null || :
+/usr/bin/gtk-update-icon-cache -qtf %{_datadir}/icons/hicolor &>/dev/null || :
+
+# Reloaded and replayed, because a rule that only takes effect at the next boot is a feature that
+# does not work on the day it is installed — and "reboot and try again" is the worst possible answer
+# to "why is the PIN option missing". The trigger is scoped to the tpmrm subsystem so this re-runs
+# rules for one device rather than for every device on the machine.
+/usr/bin/udevadm control --reload &>/dev/null || :
+/usr/bin/udevadm trigger --subsystem-match=tpmrm --action=change &>/dev/null || :
+
+%postun
+/usr/bin/update-desktop-database %{_datadir}/applications &>/dev/null || :
+/usr/bin/gtk-update-icon-cache -qtf %{_datadir}/icons/hicolor &>/dev/null || :
+
+# On the way out too, so removing the package takes the ACL with it rather than leaving the last
+# session's user holding access to the TPM until the next reboot.
+/usr/bin/udevadm control --reload &>/dev/null || :
+/usr/bin/udevadm trigger --subsystem-match=tpmrm --action=change &>/dev/null || :
+
+%files
+# The whole tree, owned by this package, which is what makes `dnf remove` take all of it. The user's
+# vault cache is not in here and is not touched by an uninstall — ClientPaths puts it under
+# $XDG_DATA_HOME, which is the Linux spelling of the same decision ADR 0013's consequences record for
+# Windows: the install moves, the profile does not, and an uninstall is not a sign-out.
+/opt/%{dodo_dir}
+%{_bindir}/%{dodo_bin}
+%{_datadir}/applications/%{dodo_bin}.desktop
+%{_datadir}/icons/hicolor/*/apps/%{dodo_bin}.png
+/usr/lib/udev/rules.d/71-dodossh-tpm.rules
+%license %{_licensedir}/%{name}/LICENSE
+
+%changelog
+# Deliberately empty. The changelog of record is the git history and the release notes on the forge,
+# and a %changelog generated per build from a version number carries no information that `rpm -q`
+# does not already answer — while a hand-written one is a third place a release has to remember to
+# edit. rpmlint says W: no-changelog and that warning is accepted.

build/macos/DodoSSH.entitlements

@@ -0,0 +1,67 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!--
+  What the hardened runtime has to be asked to relax before a .NET application will run under it.
+
+  The hardened runtime is not optional: notarization refuses a Developer ID submission without it,
+  and Gatekeeper refuses an un-notarized download. So every entitlement below is the price of being
+  distributable at all, and each one is a hole in a wall that is otherwise worth having. They are
+  listed one at a time, with what breaks without each, because the temptation when notarization
+  fails at eleven at night is to paste in a longer list from somewhere and stop thinking.
+
+  ◆ WHAT IS DELIBERATELY NOT HERE.
+
+  com.apple.security.app-sandbox. Developer ID distribution outside the App Store does not require
+  the sandbox, and turning it on would break the product outright: the terminal's data plane is a
+  loopback WebSocket (see DodoSSH.Client.Terminal/TerminalDataPlane.cs), and a sandboxed process
+  needs com.apple.security.network.server to listen at all, plus network.client to reach any host
+  the user asks for. This is the same shape of decision as ruling out MSIX on Windows, which was
+  ruled out for the same loopback reason — docs/platform-flags.md.
+
+  com.apple.security.cs.debugger. Would let this process attach to others. Nothing here debugs
+  anything, and it is the entitlement most worth not having.
+-->
+<plist version="1.0">
+<dict>
+  <!--
+    CoreCLR compiles IL to machine code at runtime and then executes the pages it just wrote. The
+    hardened runtime's default is that no page is both writable and executable, so without this the
+    process does not start — it dies during runtime initialisation, before any of this application's
+    code runs, which means before anything exists that could report it.
+  -->
+  <key>com.apple.security.cs.allow-jit</key>
+  <true/>
+
+  <!--
+    The broader form of the same permission, and it is needed as well as allow-jit rather than
+    instead of it. allow-jit covers pages mapped through the MAP_JIT convention; CoreCLR also
+    allocates executable memory outside that path — stubs, precode, and the write-xor-execute
+    fallback it uses when MAP_JIT is unavailable. With only the first, startup gets further and
+    still fails.
+  -->
+  <key>com.apple.security.cs.allow-unsigned-executable-memory</key>
+  <true/>
+
+  <!--
+    Library validation requires every loaded dylib to be signed by the same team as the main
+    binary. This bundle carries native libraries built by other people — libsodium, libSkiaSharp,
+    libHarfBuzzSharp, libe_sqlite3, libAvaloniaNative — and the release script signs each of them
+    with this Developer ID, which would in principle satisfy validation.
+
+    It is disabled anyway, and the reason is the updater. Velopack replaces the bundle in place and
+    relaunches it, and the process doing the replacing is not always signed by the same team as the
+    process being replaced during the changeover. Leaving validation on makes the failure mode of a
+    bad update "the application will not start", with no way to recover except a reinstall the user
+    would have to be told about through some other channel.
+  -->
+  <key>com.apple.security.cs.disable-library-validation</key>
+  <true/>
+
+  <!--
+    The runtime reads DYLD_ variables while resolving its own native dependencies, and Velopack's
+    update path sets them. Without this the hardened runtime strips them silently and the failure
+    surfaces later as a library that cannot be found, naming a file that is plainly present.
+  -->
+  <key>com.apple.security.cs.allow-dyld-environment-variables</key>
+  <true/>
+</dict>
+</plist>

build/macos/Info.plist.template

@@ -0,0 +1,125 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!--
+  The Info.plist for the macOS bundle, with the version left as a placeholder.
+
+  ◆ A TEMPLATE RATHER THAN A FILE, BECAUSE vpk COPIES A CUSTOM PLIST VERBATIM.
+
+  Measured, not assumed: `vpk [osx] bundle --plist` performs no substitution of any kind. It logs
+  "Bundle using provided Info.plist" and copies the bytes. That is also why it refuses --plist and
+  --bundleId together — with a plist supplied, every key is the caller's problem.
+
+  So a committed Info.plist would carry whatever version it was written with into every release
+  afterwards, and the failure is quiet in the worst way: Velopack's own release index would carry the
+  right version, the updater would compare correctly and update correctly, and only the About window,
+  Finder's Get Info panel and any crash report would claim the build was something else. Nobody
+  reads those on the day of a release. scripts/release-macos.sh substitutes @VERSION@ into a copy
+  and passes that.
+
+  ◆ WHY A CUSTOM PLIST AT ALL, WHEN vpk WRITES A PERFECTLY GOOD ONE.
+
+  Three keys it does not write, each of which is a real defect without it:
+
+    CFBundleDisplayName    The bundle on disk is DodoSSH.Desktop.app, because the pack id must not
+                           be DodoSSH — see scripts/release-macos.sh for the directory collision
+                           that rule prevents. On Windows the pack id is invisible; on macOS it
+                           names the thing in /Applications and in the Dock. This key is what puts
+                           "DodoSSH" back in front of a person while the bundle keeps the id.
+
+    LSMinimumSystemVersion Without it macOS will happily launch this on a release the runtime was
+                           never built for, and the user gets a dyld crash rather than a sentence.
+
+    NSHumanReadableCopyright  Shown in the About panel. Absent, the panel shows a blank line.
+-->
+<plist version="1.0">
+<dict>
+  <!--
+    CFBundleName is what the menu bar shows and is capped at 15 characters by convention;
+    CFBundleDisplayName is what Finder and the Dock show. Both say DodoSSH, and the bundle
+    directory does not. See the note above.
+  -->
+  <key>CFBundleName</key>
+  <string>DodoSSH</string>
+
+  <key>CFBundleDisplayName</key>
+  <string>DodoSSH</string>
+
+  <!--
+    Reverse-DNS under the domain this project actually controls. It is the identity Gatekeeper,
+    the notary service and the keychain all key off, so it is as irreversible as the Windows pack
+    id: changing it makes an update a different application, and it orphans anything the previous
+    identifier stored — including the Secure Enclave key MacDeviceKeyStore holds, which is scoped
+    to this identifier and cannot be migrated because its whole point is that it never leaves the
+    enclave.
+  -->
+  <key>CFBundleIdentifier</key>
+  <string>dev.dodotech.dodossh</string>
+
+  <!--
+    The apphost the publish produced, named for the product by <AssemblyName> in the csproj rather
+    than for the project. Must match --mainExe or the bundle launches nothing.
+  -->
+  <key>CFBundleExecutable</key>
+  <string>DodoSSH</string>
+
+  <!--
+    Both version keys take the numeric core only — 1.2.3 and never 1.2.3-rc.1 — because Apple
+    defines them as one to three dot-separated integers and notarization rejects what it cannot
+    parse. The full version, prerelease suffix and all, is in Velopack's release index, and that
+    is the one the updater compares. These two are for Finder and for Gatekeeper.
+
+    They are the same value rather than the usual marketing/build split, because there is no build
+    counter here that a release does not already bump.
+  -->
+  <key>CFBundleShortVersionString</key>
+  <string>@VERSION@</string>
+
+  <key>CFBundleVersion</key>
+  <string>@VERSION@</string>
+
+  <!--
+    The file name inside Contents/Resources, which is where --icon puts it. With a custom plist
+    nothing rewrites this key, so a rename of the asset that forgets this line produces a bundle
+    showing the generic application icon and no error anywhere.
+  -->
+  <key>CFBundleIconFile</key>
+  <string>dodossh.icns</string>
+
+  <key>CFBundlePackageType</key>
+  <string>APPL</string>
+
+  <!--
+    12.0, and it is read off the binaries rather than off a support matrix. The apphost and
+    libcoreclr.dylib in a net10.0 osx-arm64 publish both carry LC_BUILD_VERSION with minos 12.0.0,
+    so 12.0 is the oldest release these bytes are built to load on.
+
+    Microsoft's *support* statement for .NET 10 is higher than this, and that difference is
+    deliberate rather than overlooked: this key decides whether macOS refuses to launch the app at
+    all, and refusing on a release where it would in fact have run is the worse of the two errors.
+    A user on an unsupported-but-working macOS gets the application; the support matrix governs
+    what gets fixed if it misbehaves there, which is a different question.
+  -->
+  <key>LSMinimumSystemVersion</key>
+  <string>12.0</string>
+
+  <!--
+    Without this the window is drawn at 1x and scaled up, which on a Retina display turns the
+    terminal — the one surface in this application that is nothing but small text — into a blur.
+  -->
+  <key>NSHighResolutionCapable</key>
+  <true/>
+
+  <key>NSPrincipalClass</key>
+  <string>NSApplication</string>
+
+  <!--
+    False, and stated rather than left out. An agent application has no Dock icon and no menu bar;
+    this one is an ordinary windowed application and the default is already false, but the key
+    being absent is indistinguishable from somebody having removed it.
+  -->
+  <key>LSUIElement</key>
+  <false/>
+
+  <key>NSHumanReadableCopyright</key>
+  <string>© DodoTech. MIT licensed.</string>
+</dict>
+</plist>

deploy/docker-compose.dev.yml

@@ -1,7 +1,12 @@
 # Development dependencies only: PostgreSQL and Keycloak.
 #
 # The API itself runs from the IDE or `dotnet run`, so the inner loop stays fast while the
-# schema is still churning. The production stack is deploy/docker-compose.yml.
+# schema is still churning.
+#
+# There is no production compose file in this repository. This one said there was, naming a
+# deploy/docker-compose.yml that has never existed in any commit. What ships instead is the
+# container image the CI image job builds and pushes; a compose file to run it belongs to the
+# self-hosting guide in M4, and until that exists the honest thing to say is that it does not.
 #
 #   docker compose -f deploy/docker-compose.dev.yml up -d
 #   dotnet run --project src/DodoSSH.Api

deploy/keycloak/realm-dodossh.json

@@ -26,7 +26,8 @@
         "post.logout.redirect.uris": "http://127.0.0.1/*"
       },
       "redirectUris": [
-        "http://127.0.0.1/callback"
+        "http://127.0.0.1/callback",
+        "dev.dodotech.dodossh:/callback"
       ],
       "webOrigins": [],
       "protocolMappers": [

docs/adding-hosts-on-the-phone.md

@@ -1,349 +0,0 @@
-# Adding hosts on the phone
-
-The phone can read a keychain and connect through it. It cannot put anything in one. This document is the
-plan for the change that fixes that, and it is written to be picked up cold — the decisions, the reasons, the
-ordered work, and the traps that are already known.
-
-> **Status: built.** All six steps. Each one compiled with the whole suite green before the next began,
-> which is the rule the ordering below sets.
->
-> | Step | State | Notes |
-> | --- | --- | --- |
-> | 1. `HostGroupSecret` grows | **Done** | Five fields, a version rule the codec did not have, a byte pin, and the "groups are flat" prose rewritten in all four places it appeared. |
-> | 2. The `Tag` item kind | **Done** | Secret, codec, merge, cipher, repository, both registries, EF entity and the generated `AddTagItem` migration. |
-> | 3. `HostSecret` grows, `Port` goes nullable | **Done** | `TagSet`, `TagIds`, `Port` as `int?`, `AsksForPassword`, both schema versions, and `HostInheritance` — the resolver. |
-> | 4. The shared view model | **Done** | Every choke point below, plus the group editor the plan never assigned a step to and without which nothing could set a default at all. |
-> | 5. The phone | **Done** | The `+`, the sheet, both editor cards, the back guard and EDIT on the connect bar. |
-> | 6. Tests and false prose | **Done** | Plus a fourth "groups are flat" site in `README.md` this plan did not list, and a cross-vault resolution bug the scouting for it turned up. |
->
-> **Tags shipped after the six steps, in a seventh pass.** They were storable and unreachable when step 6
-> closed — a full item kind that no screen drew. Both heads now show a chip per tag on a host row and toggle
-> them in the host editor, which also creates one inline, because wanting a tag and tagging a host are the
-> same moment. Renaming and deleting live in a TAGS category on the keychain screen: a tag is an item so
-> that renaming it is one write instead of twenty, and that rename needed somewhere to happen.
->
-> That pass also made the desktop host editor scroll. A picker's height is a chip per tag in the keychain,
-> so no fixed height holds it — which the layout suite caught the moment its seeder grew tags.
->
-> **`HostFields.From` was answered by a refusal rather than by threading the resolver into the sync engine.**
-> A relay host may not inherit its port; `HostSecret.TryValidate` refuses one that tries. A plaintext column
-> derived from a *different* item goes stale when that item is edited and nothing re-pushes the hosts beneath
-> it, so the relay would keep dialling the old port — which is worse than the restriction.
->
-> **One decision was taken that this plan did not specify.** "Three states where there were two" is four, not
-> three: a host can bind a key, bind a credential, be pinned to a typed password, or take its group's answer,
-> and two nullable ids express three of those. Naming neither id now means *inherit*, so
-> `HostSecret.AsksForPassword` was added to say "a typed password, even under a group that lends a key" out
-> loud. Nothing stored changed meaning — no group could lend a binding before this build, so every existing
-> host resolves exactly as it did.
-
-## What was asked for
-
-A floating `+` at the bottom right of the phone's HOSTS screen, opening a menu at the bottom offering **new
-host** or **new group**.
-
-A **host** carries: which vault it is filed into (personal or a team's), an alias, a hostname or IP address, a
-group, tags, an SSH port, and a username with either a password or an SSH key.
-
-A **group** carries: an alias, an optional parent group, and a default SSH port, username and
-password-or-key — **defaults for the hosts inside it**.
-
-## The shape of the problem
-
-Three of those fields do not exist at any layer, and one of them is refused on the record.
-
-| Asked for | What exists today |
-| --- | --- |
-| The `+` and the editors behind it | Nothing. `Theme/Phone.axaml` has no `.fab` class, and its comment says an unused style would be "a claim that the control exists somewhere". `HostsScreen.axaml`'s own v2 note says hosts are created on the desktop and sync down. *(This row named a `ConnectionsScreen.axaml` as a second site. No such file exists or ever has; and the `HostsScreen` statement is an XAML comment, so no phone screen ever rendered that sentence to a user.)* |
-| Vault picker | **Built.** `VaultViewModel.TargetVaults` / `SelectedTargetVault` / `HasVaultChoice`, hidden at one vault. The desktop's keychain screen already draws it. |
-| Alias, hostname, port, username, key-or-password, group | **Built**, in the shared `VaultViewModel` host editor — `EditorLabel`, `EditorHostname`, `EditorPort`, `EditorUsername`, `EditorAuthenticationChoices`, `EditorGroupChoices`, `SaveHostCommand`. The phone has never bound any of it. |
-| Tags | **Nothing.** `SyncEntityType.Tag = 5` and `HostTag = 6` are reserved slots with nothing behind them. `HostSecret` has no tag field. |
-| A group's parent | **Refused on the record.** `HostGroupSecret`'s own remark says groups are flat because two clients can each re-parent A under B and B under A offline, a scalar merge accepts both, and the result is a cycle no reader can draw and the server cannot see, because it is inside the payload. |
-| A group's defaults | **Nothing.** `HostGroupSecret` has exactly one field, `Label`. |
-
-So the phone half is mostly wiring. The domain half is three new capabilities, and one of them overturns a
-written decision.
-
-## The three decisions
-
-**Tags are a real vault item, and a host names them.** A tag is a shared, reusable thing — the point of it is
-to put the same tag on twenty machines and filter by it later — so it needs an identity of its own, not a
-string repeated inside twenty payloads. `SyncEntityType.Tag = 5` becomes a live item kind. Membership is a
-`TagIds` set on `HostSecret`, **not** the reserved `HostTag = 6` join.
-
-The reason to skip the join is that the merge does not need it. The argument for `HostSecret.GroupId` living on
-the host is that filing two hosts into one group must be two writes to two items; the same holds here. What a
-join would have bought beyond that is two machines tagging *the same* host without one of them losing — and
-`ThreeWayMerge.Map` already gives that. It resolves a keyed collection key by key, honouring presence and
-absence, which is set semantics with removals. `HostOptions` is merged through it today via `ToNameMap()`, and
-a tag set keyed by tag id is the same shape. So `HostTag = 6` stays reserved and unused, and the second run
-through the item-kind checklist is not spent.
-
-**Group defaults are inherited, not copied.** A host that leaves a field blank uses its group's value; a host
-that fills one in overrides it. The editor shows the group's value as the field's placeholder, so the form
-says what leaving it blank will get you. Copying the values into the host at create time was the alternative
-and it was rejected: it makes a group a one-shot template rather than a live default, and changing a group
-would then leave every host that had ever been created under it pinned to the old value.
-
-**Groups get a parent, and the resolver breaks cycles.** `HostGroupSecret` gains `ParentId`. The editor refuses
-a parent that is already a descendant, which stops a cycle being made here; the resolver walks with a visited
-set and stops at a repeat, which contains a cycle that arrives from somewhere else. A cycle therefore degrades
-to a group that reads as a root — flat headings, defaults unresolved past that point — instead of hanging the
-connect path, and clearing the parent in the editor is the repair. **This is the load-bearing part of the
-decision:** with inheritance the chain is walked at connect time, so an unguarded cycle is not an undrawable
-sidebar, it is a shell that never opens.
-
-`HostGroupSecret`'s "No parent. Groups are flat." remark and `HostGroupSecretTests`' class remark both become
-false prose and must be rewritten to say what replaced the argument, not deleted. Same for the four rows in
-`docs/design-import-gaps.md` recording the phone's missing `+`.
-
-## What the payload becomes
-
-```
-HostSecret                            HostGroupSecret
-  Label                                 Label
-  Hostname                              ParentId            ← new, optional
-  Port            int? — null inherits  DefaultPort         ← new
-  Username        null inherits, "" none DefaultUsername    ← new
-  Notes                                 DefaultSshKeyId     ← new
-  JumpHostIds                           DefaultCredentialId ← new
-  Options
-  SshKeyId        null inherits       TagSecret
-  CredentialId    null inherits         Label
-  AsksForPassword ← new, true only
-  GroupId         null = ungrouped
-  TagIds          ← new
-  RelayEnabled
-```
-
-`Port` becoming `int?` is the change with the widest blast radius, and it is unavoidable: `int` with a default
-of 22 has no way to say "I have no port of my own". `HostSecret.DefaultPort = 22` stays, as the last fallback
-after the chain runs out.
-
-### Three states where there were two
-
-`SshKeyId` and `CredentialId` both null currently *means* "ask for a password each time" — a decision, not an
-absence. `AuthenticationChoice`'s own remark argues at length that the two must never be conflated.
-Inheritance adds a state, so the picker needs an explicit **"Inherit from group"** entry beside **"Password
-(ask each time)"**, and `Bound(...)` needs to distinguish them. `Username` has the same problem: null means
-"no username" today and is refused at connect; it has to come to mean "inherit", with "no username" still
-reachable and still refused.
-
-**Built as four states, not three, because two nullable ids only carry three.** Key, credential, typed
-password, inherit — and naming neither id was the third and is now the fourth. `HostSecret.AsksForPassword`
-carries the difference: null is "not stated", which walks the chain and lands on a typed password if the
-chain lends nothing, and `true` is "a typed password, even under a group that lends a key". Only `true` is
-ever written, and a decoded `false` folds back to null, so a host that never touched the field encodes
-exactly as it did before it existed. `Username` needed no field — an empty string is "no username" and null
-is "inherit". `Port` needed none either: there is no "explicitly no port", only 22 at the end of the chain.
-
-Mutual exclusion moves with it. `HostSecret.TryValidate` enforces "a key or a credential, never both" per
-record; a host naming a credential under a group naming a key is two individually valid records that resolve to
-two bindings. **Exclusion has to be enforced at resolution**, host-over-group, and the resolver must never
-return both.
-
-## The order to build it in
-
-Each step compiles and its tests pass before the next one starts. Steps 1–3 are the domain and are
-prerequisites for everything; step 4 is the only user-visible one.
-
-### 1. `HostGroupSecret` grows
-
-`src/DodoSSH.Client.Domain/HostGroupSecret.cs`, `HostGroupSecretCodec.cs`, `HostGroupSecretMerge.cs`.
-
-- Add `ParentId`, `DefaultPort`, `DefaultUsername`, `DefaultSshKeyId`, `DefaultCredentialId`.
-- `TryValidate`: reject `Guid.Empty` for each id, reject a `ParentId` equal to the group's own id where that is
-  knowable, apply the same 1–65535 range rule to `DefaultPort`, and refuse a group naming both a default key
-  and a default credential.
-- **The codec needs the host codec's rule, which it does not have today.** `HostGroupSecretCodec` stamps
-  `CurrentSchemaVersion` unconditionally. Copy `HostSecretCodec.SchemaVersionFor` — a *maximum over the fields
-  actually present*, not a ladder — or upgrading one machine makes every group read-only on every other. A
-  group carrying none of the new fields must still encode at version 1, byte for byte.
-- `HostGroupSecretMerge` gains the five fields, each a scalar.
-- Add a group byte pin mirroring `AddingTheKeyField_DidNotChangeTheBytesOfAHostWithoutOne`. The group codec has
-  none, and the same "old vaults must re-encode identically" argument applies to it.
-
-### 2. The `Tag` item kind
-
-`SyncEntityType.Tag = 5`, `CryptoSpec.AadResourceType.Tag = 8`, `ChangeEntityType.Tag = 5` and the
-`PublicAPI.Unshipped.txt` line **already exist** — no contract change, no crypto-spec change, no `docs/crypto.md`
-change. `LocalCacheProtector`'s AAD switch already maps `Tag`.
-
-New: `TagSecret.cs`, `TagSecretCodec.cs`, `TagSecretMerge.cs` in `DodoSSH.Client.Domain`; `TagCipher.cs`,
-`TagRepository.cs` in `DodoSSH.Client.Sync`. Edited: `Client.Sync/ItemKinds.cs` (registry entry + `TagKind`),
-`VaultSession.cs` (a `Tags` repository beside `HostGroups`), `Domain/Hosts.cs` (`VaultTag`),
-`Infrastructure/DodoDbContext.cs`, `Infrastructure/Configurations/HostAndSyncConfigurations.cs`,
-`Api/Features/Sync/ItemKinds.cs` (registry entry + server `TagKind`).
-
-**A server migration is required** — a tag is its own table. Generate it, do not hand-write it:
-`dotnet ef migrations add AddTagItem --project src/DodoSSH.Infrastructure`, committing all three resulting file
-changes. **No client migration**: `ClientCacheContext` has one `item` table keyed by `(VaultId, EntityType,
-EntityId)`, so a new kind is new rows.
-
-Four traps, all of which pass a naive test suite:
-
-- **`TagCipher` must say `AadResourceType.Tag` as a constant, never cast from `SyncEntityType.Tag`.** Tag is 5
-  on the wire and 8 in the crypto enum. A cast seals tags under `Credential`, which encrypts and decrypts
-  perfectly on the machine that wrote it and is a spec violation nothing notices until another implementation
-  refuses the item — and the AAD is frozen into stored ciphertext. `AadResourceTypeTests` is the test that
-  catches this, and it only catches it once `PinnedPairs` and `SealSample` have a Tag arm.
-- **The `ItemKinds.Registry` entry is what `SyncEngine` pulls with.** `SyncedTypes` is projected from it. Miss
-  it and tags encrypt, merge and list perfectly on the machine that made them and are never once requested from
-  the server.
-- **`TagKind.Fields` returns `null`, not an empty `SyncPlaintextFields`.** An empty record still serialises
-  `relayEnabled: false`, which invites a reader to believe the type has a relay setting that is off.
-- **`ValidateFields` must refuse everything, with reasons.** There is no default; an unimplemented one accepts
-  whatever arrives. `RelatedId` matters more here than it did for a group — it is the field a future `HostTag`
-  would reach for, and `SyncPlaintextFields` is frozen so it cannot be removed. Follow `SnippetKind` and
-  `ObjectStoreKind`, which are more complete than `HostGroupKind`.
-
-### 3. `HostSecret` grows, and `Port` goes nullable
-
-`HostSecret.cs`, `HostSecretCodec.cs`, `HostSecretMerge.cs`, plus a new `TagSet` value type and a new resolver.
-
-- `TagSet` is a structural-equality wrapper in the shape of `JumpChain` — a plain `IReadOnlyList<Guid>` on a
-  record gets reference equality from the compiler-generated `Equals`, which would make every host read as
-  changed on every sync pass and two identical edits register as a conflict. Unlike `JumpChain` it is a *set*:
-  order carries no meaning, so it sorts and dedupes, and it exposes a `ToIdMap()` for the merge.
-- `HostSecretMerge` merges `TagIds` through `ThreeWayMerge.Map`, per tag, so two people each adding a different
-  tag to one host both keep theirs. It is not a scalar; a whole-value merge here would silently drop one side.
-- `Port` becomes `int?`. `TryValidate`'s `is < 1 or > 65535` is already false for null, which is the wanted
-  behaviour but is silent — say so in a comment, and fix the message, which interpolates `{Port}` and renders
-  empty for null.
-- **`SchemaVersionFor` needs branches for both new fields, and this is the one that loses data if forgotten.**
-  A host with `Port = null` written at version 1 omits the property (`WhenWritingNull`); an older client
-  deserialises `int Port` as 0, `TryValidate` refuses it, and the item reads as **unreadable rather than
-  read-only** — gone, not locked. Add `PortInheritSchemaVersion = 5` and `TagIdsSchemaVersion = 5` (or 5 and 6)
-  and make both bump the maximum.
-- **`AddingTheKeyField_DidNotChangeTheBytesOfAHostWithoutOne` pins the exact bytes** of a minimal host,
-  including `"port":22`. A host that now inherits its port drops that key. Keep the pinned literal for a host
-  with an explicit port, and add a companion for the inheriting one — do not edit the old literal to match
-  whatever the code produces.
-- `TagIds` must serialise **last** in `HostPayloadDocument`, and empty must serialise as absent, or every host
-  in every vault looks changed on the first sync after the upgrade.
-
-New file: the resolver. Given a host and the group list, walk `GroupId` then `ParentId` with a visited set,
-returning resolved port, username and binding, plus which group each came from — the editor needs the second
-half for its placeholders. It must never return both a key and a credential.
-
-### 4. The shared view model, and the two heads
-
-`VaultViewModel` is where nearly all of it lands, and both heads get it at once because both bind the same
-properties.
-
-The choke points, from the connect-path trace:
-
-- **`TryBuildAuthentication(HostSecret, …)` and `TryBuildConnectionRequest(HostSecret, …)`** both take a bare
-  `HostSecret` — this signature is where group context is currently lost. Both must take the resolved value.
-  This is the single highest-leverage change: it is the only auth resolution in the product, and both heads
-  and both transports go through it.
-- The credential-username fallback becomes three levels: credential → host → group chain.
-- `Complete(...)` must refuse on an empty username only *after* the chain has been consulted.
-- `HostRowViewModel.Address`, `Dialled`, `TransfersViewModel`'s connected-to label and
-  `TerminalWorkspace.Describe` must show the **dialled** address. `MainWindowViewModel.Rank` searches
-  `Address`, so an unresolved one makes hosts unfindable by their real port.
-- **`HostFields.From` in `HostCipher` must emit the resolved port.** This is the one place where an unresolved
-  read is a wrong wire rather than a wrong label: it fills the relay's plaintext columns, which is what the
-  *server* dials. A relay host with a null port also fails server validation outright.
-- Host-key pinning keys on host and port. `ForgetHostKeyAsync` and the "pin in use" endpoint set read
-  `row.Host.Port` directly — a host inheriting 2222 pins under 2222 while both look under 22.
-- `HostRowViewModel.Authentication`, `SelectedHostAsksForAPassword` (in **both** `VaultViewModel` and
-  `TransfersViewModel`) and `SelectedHostAuthenticationNote` all switch on the host's own two ids and must
-  switch on the resolved binding, with new wording for a binding that came from a group.
-- **`HostsBoundTo`** counts hosts naming a key explicitly. A key deleted while only a *group* names it warns
-  nobody and then refuses every host under that group at connect time. It must walk groups too.
-- The editor: load must distinguish "unset" from "explicitly this", `BuildHost` must preserve that in both
-  directions, and each field needs the resolved value as its placeholder.
-- `ImportViewModel.IsAlreadyPresent` compares port and username directly, so a stored host inheriting 22 will
-  never match an imported host at 22 and the import screen re-offers duplicates. Decide whether an imported
-  block with no `Port` should inherit rather than pin 22 — `SshConfigResolver` returns `DefaultPort` today.
-
-### 5. The phone
-
-`Views/HostsScreen.axaml` and `Theme/Phone.axaml`. No csproj edit — Avalonia globs `**/*.axaml`. Compiled
-bindings are on, so a binding to a property that does not exist is a build error; `TreatWarningsAsErrors` is on
-repo-wide, so any XAML warning fails CI.
-
-- A `Button.fab` style, and the comment that currently refuses one rewritten to say what it now opens.
-- The `+` needs a `Panel` wrapping the screen's root `Grid` so it can sit bottom-right without moving. Placing
-  it inside the list row would put it above the connect bar, which appears and disappears with the selection.
-- The bottom sheet follows `HostKeySheet.axaml`, the head's one existing sheet: a scrim
-  (`Border Background="#9E0E1220"`, a literal because the palette holds no alpha surface), a bottom-aligned
-  `Border` with `CornerRadius="22,22,0,0"`, a decorative grab handle, and **no tap-to-dismiss** unless it is
-  wired by hand, because nothing on this head implements it.
-- **Scrim reach is a real choice.** Declared inside `HostsScreen`, a scrim darkens the screen area only and the
-  64px nav bar stays lit and tappable. To cover the display the sheet has to sit in `PhoneShell.axaml` beside
-  `HostKeySheet`, which is exactly why that control lives there.
-- **The back gesture needs a new guard**, above the `switch` in `PhoneShell.axaml.cs`, in the shape of the
-  host-key guard: back should close the sheet or the editor rather than leave the screen. The comment on that
-  method states an invariant about the switch cases staying in step with the surface tests — so add a guard,
-  do not extend the switch. *(This named `IsVaultsSurface` and `IsConnectionsSurface`. Neither exists
-  anywhere in the repo; the real properties are `IsMoreSurface` and `IsTerminalSurface`, and only the first
-  is what the switch mirrors.)*
-- The editors follow `SnippetsScreen.axaml`: a `Border Classes="card"` in the same grid row as the list,
-  toggled on `IsEditing`, with the list toggled on `!IsEditing`. There is no dialog, no modal and no
-  editor-screen anywhere on this head, and the reason is written down — a form stacked over the list hides what
-  it is about.
-- Editing an existing host comes nearly free: the shared `IsEditing` panel serves both, so the connect bar gets
-  an EDIT button beside CONNECT. Not asked for, but a `+` that adds hosts with no way to correct one is a
-  strange thing to ship.
-
-### 6. Tests
-
-House style: `Subject_Predicate`, articles included, the predicate smuggling in the reason
-(`_RefusesRatherThanFallingBackToThePassword`). No `Async` suffix on test methods; always on private helpers.
-Every class carries a remark naming the class of failure it exists to catch. xunit.v3 + Shouldly, with the
-because-string used as prose.
-
-- Three shared builders need the new fields: `Domain.Tests/HostFactory.cs`, `Sync.Tests/SyncHarness.cs`,
-  `Sync.Tests/HostCipherTests.cs`. There is no shared group factory and with five new fields there should be.
-- Three guard tests fail by design and that is what they are for:
-  `HostSecretMergeTests.EveryScalarField_IsRoutedThroughAMerge` (note `GroupId` is missing from its `with`
-  block today — an existing gap worth closing while there) *(closed: `GroupId` is in that block now)*,
-  `ValueSemanticsTests.TryValidate_RejectsWhatCannotBeStored`, and the two "groups are flat" remarks.
-- `ItemKindsTests.ThePullFilterNamesEveryTypeThisBuildSynchronises` is an ordered list — add Tag in registry
-  order. `AadResourceTypeTests` needs `PinnedPairs`, a `SealSample` arm and a `NewTag()` builder.
-- `SyncEndpointTests.AHostGroupCarryingAParent_IsRejected` asserts the server refuses a plaintext `ParentId`
-  with a reason containing "flat". Putting the parent **inside the payload** does not break it mechanically,
-  but its stated reason becomes wrong. It should survive with a rewritten comment — the parent lives in the
-  payload, the plaintext column stays refused, and ADR 0004 is why. Decide that deliberately rather than
-  letting it drift. *(Decided that way, and the name says it now:
-  `SyncEndpointTests.AHostGroupCarryingAPlaintextParent_IsRejected`. The column is still refused and the
-  reason is the trust model rather than flatness.)*
-- `Push_AnUnsupportedEntityType_IsInvalidNotAFailedBatch` picks the first type with no kind registered, very
-  likely Tag today. It self-heals to `HostTag` via `Assert.SkipWhen`.
-- `App.Layout.Tests` measures pixels and bends rather than breaks. Extend the seeders — nesting and a tag row
-  change what the widest realistic content is — rather than the assertions. `ScreenLayoutTests.cs:1310` pins
-  `ChangedFields = "Hostname, Port, Username, Options, Group"` as a literal; add `Tags` to keep it measuring
-  the worst case.
-- `ShellFlowTests.ReadyToConnectAsync` gives its host both a username and a port, so the inheritance tests need
-  a host with neither, or the group is never the source.
-
-## Found on the way, and out of scope
-
-`LocalCacheProtector`'s AAD switch is missing `ConnectionLogEntry`, `ActivityLogEntry` and `ObjectStore`, so
-`ConflictStore.Record` throws `ArgumentOutOfRangeException` on a conflict for any of those three. Pre-existing,
-unrelated to any of this, and `Tag` is already in that switch. Worth a separate fix.
-
-> **Fixed 2026-08-04**, having outlived the phases that shipped the logs and the buckets — which is exactly
-> the drift a note like this is meant to prevent, so it is worth saying what let it last. Of the three callers
-> of `AadResourceTypes.For`, two reach it only when an item carries plaintext fields and none of these three
-> does; the third, `ConflictStore.RecordAsync`, calls it unconditionally but is reached only by a real merge
-> conflict, which every existing test raised against a `Host`. The arms are in, and two tests now hold them
-> there: `CacheStoreTests.AConflict_CanBeRecordedForEveryKindOfItem` records one per kind, and
-> `AadResourceTypeTests.EverySyncableType_HasAnArmInTheStorageMapping` fails on the *next* item type added
-> without one, by name rather than by a list kept by hand.
-
-## Prose that becomes false
-
-Not a tidy-up — these are the places the codebase currently tells a user or a maintainer that this feature is
-deliberately absent, and each states a reason that will no longer hold.
-
-- `Theme/Phone.axaml` — the "No floating action button" block.
-- `Views/HostsScreen.axaml` — the v2 and v3 notes on the missing `+`.
-- ~~`Views/ConnectionsScreen.axaml`~~ — **this file does not exist and never did.** The real inventory is
-  the two above plus `README.md`, `docs/android-port.md` and `docs/design-import-gaps.md`. Neither phone
-  site was "rendered on screen": both are XAML comments.
-- `HostGroupSecret.cs` and `HostGroupSecretTests.cs` — "No parent. Groups are flat."
-- `docs/design-import-gaps.md` — the **Add host** row, the two floating-button rows, the "Tag / HostTag still
-  reserved and unused" paragraph, and the tag-chips and groups-on-a-team's-hosts rows in the Hosts table.

docs/adr/0004-relay-authorization.md

@@ -99,3 +99,22 @@ real ones.
 - **Client-supplied target address.** One line of convenience, and the relay becomes an open
   proxy into the operator's network.
 - **Allow-listing by hostname.** Defeated by DNS rebinding.
+
+## Implementation notes
+
+Built in `DodoSSH.Api/Features/Relay` and, on the client, `RelayBridge` and `ServerRelay`; see
+`docs/reaching-a-host-you-cannot-dial.md`, "What step 3 built". Three places differ from the text above:
+
+- **Backpressure is an awaited copy loop, not `System.IO.Pipelines`.** Each direction reads one buffer and
+  writes it before reading again. The property is the same — a slow peer stops the read, and TCP throttles
+  the origin — and the bound is tighter: two 32 KiB buffers per session.
+- **There is no `?ticket=` fallback.** Both clients can set a subprotocol, and the fallback's cost — a bearer
+  credential in the URL — was the reason it needed scrubbing at all. It can be added if a constrained
+  client ever needs it; the scrubbing requirement stands.
+- **The ticket carries no device id.** The server does not see a device on an ordinary request, so there was
+  nothing to put there. `PortForwardId` on the request is refused: a forward's port is ciphertext, and
+  honouring one would mean taking the port from the client.
+
+The loopback bridge serves the relay only. SSH.NET's `ForwardedPortDynamic` already is the jump-host path's
+listener, so "one mechanism, two features" is true one level down: both features reach their target through
+a loopback SOCKS5 proxy named on the connection request.

docs/adr/0007-device-key-protection.md

@@ -1,4 +1,4 @@
-# ADR 0007 — What protects the device key on Windows
+# ADR 0007 — What protects the device key on the desktop
 
 **Status:** accepted, 2026-07-30
 **Supersedes nothing. Constrains** the device-unlock work described in the client roadmap.
@@ -107,6 +107,81 @@ almost nil. Two things to check when it is taken up: `EnrollmentValidation` pins
 `DevicePublicKey` to `CryptoSpec.PublicKeySize` (32 bytes; a P-256 public key is 33 or 65), and the
 envelope's minimum-length rule.
 
+### 4. Linux: sealed into the TPM behind a PIN the TPM checks
+
+Added after the Linux head was packaged ([ADR 0013](0013-desktop-distribution-and-updates.md) decision 11).
+The third desktop platform reached this file the way the first two did — by having to answer the amendment
+above — and it is the only one where the answer is not a clean yes.
+
+**The mechanism.** The 32-byte device key is sealed into a TPM 2.0 keyedhash object whose `userAuth` is
+derived from a PIN. Unsealing is a command the TPM refuses without it. The blob on disk is encrypted to the
+machine's storage hierarchy, so it is useless on any other machine, and nothing in this application ever
+compares the PIN. The parent is an ECC P-256 primary rederived from a fixed template on every use rather
+than made persistent — a persistent handle would occupy an NV slot forever, need cleaning up on withdrawal,
+and collide with whatever else on the machine chose the same index. ECC rather than RSA because creating an
+RSA primary on a discrete TPM takes seconds and this runs on the unlock path.
+
+**The sealed object deliberately does not carry `NoDA`, and that single attribute is what makes a
+four-digit PIN defensible.** Every wrong answer advances the TPM's dictionary-attack counter and the chip
+stops answering entirely once it trips — measured against a software TPM at four wrong tries, after which
+the *correct* PIN gets nothing either. That is the property a short secret has here and cannot have in a
+file, and it is asserted by a test rather than left to a reading of the template, because removing the
+attribute would leave every other assertion in that file passing.
+
+◆ **This is weaker than Windows Hello and Touch ID, and the difference is invisible to the user.** Both of
+those raise their prompt from inside the platform — a secure desktop, a system sheet — so the process asking
+never sees the secret. Linux offers nothing to borrow, so the PIN is typed into a window this application
+draws, and an attacker already running as the user could draw a convincing copy of it. Two things bound
+that, and neither erases it: the same attacker could counterfeit the passphrase screen, which is a much
+larger prize, so the feature creates no capability that did not exist; and what they would obtain is a PIN
+that is worthless without this machine's TPM, where the passphrase is worth every machine. **The parity is
+not claimed anywhere in the interface**, which is why `DesktopDeviceKeyStores` documents the ordering and
+why the prompt is deliberately plain rather than dressed to look like a system dialog — teaching people
+that a convincing DodoSSH PIN box is a system dialog is the one thing this design must not do.
+
+**Reaching the TPM at all is the other half, and the package now arranges it.** On Fedora `/dev/tpmrm0` is
+`root:tss` mode 0660 and desktop users are not in `tss`, so without help the ordinary case is a working TPM
+this process may not open — which reports unavailable and keeps asking for the passphrase, exactly as a
+machine with no TPM does.
+
+◆ **This originally said shipping a udev rule to widen that was refused, on the grounds that it is a
+system-wide decision an application package should not make. That refusal was overturned deliberately, and
+the reasoning it was replaced with is narrower rather than louder.** The refusal treated "widen access" as
+one thing. It is two, and they are not close:
+
+- *Adding a person to `tss`* is what the refusal was really about, and it remains the wrong answer. It is
+  permanent, it applies whether or not they are at the machine, it applies over SSH, and it needs an
+  administrator and a fresh login before the feature appears at all.
+- *Tagging `/dev/tpmrm0` with `uaccess`* is the mechanism systemd already uses for `/dev/dri`, webcams,
+  scanners and FIDO keys: logind grants a POSIX ACL to the user of the **active local session** and removes
+  it when that session ends. Not over SSH, not on an inactive VT, not to daemons. `build/linux/71-dodossh-tpm.rules`
+  is that one line, and it changes neither the mode nor the group, so anything that worked before still does.
+
+**`tpmrm` and never `tpm`, which is what makes it defensible at all.** `/dev/tpm0` is the raw device and is
+exclusive; a process holding it can leave the chip in a state that breaks every other user. `/dev/tpmrm0` is
+the kernel's resource manager, which virtualises handles and isolates callers — it exists *because* the TPM
+is meant to be usable by more than one unprivileged program. Granting a seat's user the manager is the use
+it was designed for.
+
+**The cost is a denial of service and not a disclosure, and it is stated because it is easy to miss.** TPM
+objects are protected by their own auth values, so this lets nobody read anybody else's sealed data. What it
+does allow is *failed* authorisation attempts — and the chip answers those by counting them and eventually
+refusing everyone, which is the same lockout decision 4 relies on. On a machine whose disk is unlocked by
+the TPM at boot (`systemd-cryptenroll`, Clevis), somebody sitting at it can trip that lockout and stop the
+next boot from unlocking until it recovers. An administrator who does not accept that masks the file with
+`ln -s /dev/null /etc/udev/rules.d/71-dodossh-tpm.rules`, which is the standard override, survives upgrades,
+and puts the behaviour back to the passphrase-only path. The rule says so in its own comments, because the
+person who needs to know is reading the file rather than this document.
+
+**What was rejected, and why each would have failed the amendment above.** The Secret Service API —
+gnome-keyring or KWallet — is out of process and OS-managed and has *no per-use consent*: once the session
+keyring is unlocked anything running as the user reads the key silently, which is DPAPI's shape exactly and
+what "Why not DPAPI alone" already refuses. `fprintd` is an authentication oracle rather than a key store:
+it answers a question and holds nothing, so skipping the call skips the gate — the amendment's mistake with
+a fingerprint reader attached. A polkit-gated root helper would genuinely work and was declined on surface
+area: it puts a privileged component in a package that installs nothing setuid, to reach a property the TPM
+already provides.
+
 ## Consequences
 
 ### The cache key had to move, and the spec changed
@@ -138,9 +213,19 @@ would have become false under DPAPI alone. A gesture is still something the atta
 
 ### Operational
 
-- **A TPM is not always there.** A machine without one gets a store that reports itself unavailable, so
-  unlock keeps asking for the passphrase and neither affordance appears in the interface. The passphrase path
-  is therefore required, not a nicety.
+- **A TPM is not always there, and on Linux it is often there and out of reach.** A machine without one gets
+  a store that reports itself unavailable, so unlock keeps asking for the passphrase and neither affordance
+  appears in the interface. The passphrase path is therefore required, not a nicety — and on Linux it is the
+  default rather than the exception, because the device node is not readable by an ordinary desktop user
+  until somebody says it should be. See decision 4.
+- **macOS reaches the same decision through different hardware, and the argument transfers intact.**
+  `MacDeviceKeyStore` puts the wrapping key in the Secure Enclave under an access control requiring user
+  presence, so Touch ID or the login password is a condition of *using* it and the enforcement is the
+  platform's rather than the process's — which is the entire point of the 2026-07-30 amendment above, and
+  the thing a self-drawn prompt over a protected file would fail to be. The mechanical differences are
+  incidental: P-256 with ECIES because the enclave holds no other kind of key, and no prompt when sealing
+  because the public half needs no consent. See docs/platform-flags.md for the three ordinary Macs where the
+  probe answers no, one of which is every unsigned development build.
 - **The stored key must be treated as losable at any time** — a reset PIN, a cleared TPM, a replaced key.
   Every loss degrades to a passphrase prompt and never to a locked-out vault, which is why every failure in
   the store returns null rather than throwing and why the three unlock statuses all end in the same advice.

docs/adr/0013-desktop-distribution-and-updates.md

@@ -228,6 +228,194 @@ changes.
 token, and it puts a compellable third party in the signing path — which is ADR 0011 rule 3's shape one
 layer down, declined there for reasons that do not stop applying because the vendor changed.
 
+**This rule is Windows-only, and macOS gets the opposite one.** See decision 10: there is no "unsigned for
+now" available on that platform at any price, because Gatekeeper refuses rather than warns.
+
+### 10. macOS is a second desktop platform on the same machinery, signed from the start
+
+The macOS head is the same application, the same Velopack, and the same two-phase person-run release. Four
+things differ, and each is forced rather than chosen.
+
+**Signing is a precondition, not an improvement.** Decision 8's whole argument — one dialog per user per
+lifetime, buy a certificate when a stranger is invited to install — has no macOS equivalent. An
+un-notarized download is refused outright, so the Developer ID certificate and the notarization round trip
+are the price of the package existing. `scripts/release-macos.sh` therefore refuses to run without the
+signing identities, where the Windows script refuses nothing.
+
+**The channels are `osx` and `osx-nightly`, and they are separate for decision 9's reason.** Four channels
+now publish to one repository, and the only thing keeping a Mac from being offered a Windows package is
+that it never reads that index. The macOS nightly channel is named and has no publisher: CI builds and
+bundles the macOS head to prove it still builds, and uploads nothing, exactly as it does for the Windows
+release channel.
+
+**The pack id is shared with Windows, and on macOS it is visible.** vpk names the bundle after the pack id,
+so `/Applications` holds `DodoSSH.Desktop.app`. Decision 2's reasoning applies with more force here rather
+than less: a pack id of `DodoSSH` would put Velopack's install root on `~/Library/Application
+Support/DodoSSH`, which is `ClientPaths.DataDirectory`, and an uninstall would take the user's un-synced
+outbox with it. `CFBundleDisplayName` puts the product name back in front of a person; the directory keeps
+the id.
+
+**arm64 only, because the check is the scarce thing.** Velopack keys a channel to one architecture, and an
+Intel package would be the only artefact in this repository reaching users without somebody having walked
+Phase 18 against it. The engineering for a second channel is small and is described in the release script;
+what is missing is an Intel Mac to verify on, and shipping blind is the thing this project's manual-check
+discipline exists to refuse.
+
+**And one thing that does not differ, which is worth saying because it is the expensive half.** The
+capability to publish still lives on a person's machine and never in CI. Notarization does not change that:
+Apple's ticket says this build came from this developer account, and says nothing about whether the build
+should have been made. Velopack clients still apply what their feed serves. Rule 3 is untouched.
+
+### 9. There is a second desktop channel, published by CI, and it is a second application
+
+[ADR 0014](0014-android-updates.md) gave the phone a nightly channel and rule 3 above gives the desktop
+none, which left the two heads with different answers to the same question — *how does somebody try what
+is on main?* — for no reason other than the order the work happened in. This is the desktop's answer, and
+it is the phone's arrangement with one difference that changes how much of it has to be built.
+
+**Android gets the separation from the platform. Windows does not.** Two Android channels cannot replace
+one another because the installer refuses a package signed by a different key; a mistake there is loud.
+Velopack applies what its feed serves and verifies no signature, so on this head the separation is entirely
+construction:
+
+| | release | nightly |
+| --- | --- | --- |
+| pack id | `DodoSSH.Desktop` | `DodoSSH.Desktop.Nightly` |
+| Velopack channel | `win` | `win-nightly` |
+| feed | newest non-prerelease release | the rolling `nightly-desktop` prerelease |
+| profile | `%LOCALAPPDATA%\DodoSSH` | `%LOCALAPPDATA%\DodoSSH.Nightly` |
+| cut by | a person, from a `v*` tag | CI, on every push to main |
+
+Four separations rather than one, because each closes a different door. The pack id decides the install
+directory and what an installed client matches an update against, so it is what makes a nightly install
+*beside* rather than *over*. The channel names the index file on the feed, so neither build ever reads the
+other's — and the release channel additionally refuses prereleases, which is belt and braces in the one
+direction that matters: an unsigned CI build must never reach a machine somebody is trusting with their
+credentials. The profile is the one that is easy to skip and would hurt most: the cache schema is migrated
+on every launch, before unlock, so a shared profile means a nightly quietly upgrading a database the
+release build then opens.
+
+**The prerelease flag is also how the two heads stay out of each other's way.** The phone's release channel
+reads `releases/latest`, which skips prereleases. A desktop nightly published as a stable release would
+become the newest release in this repository, and every phone on the release channel would start failing
+its check against a release carrying no Android manifest. The nightly is published with `--pre` for its own
+sake and for that one.
+
+**What this costs, stated where somebody will read it before installing:** whoever can write a release on
+this repository can put a build on every nightly desktop, because the client applies what the feed serves
+without verifying a signature. That set includes CI and therefore everyone who can change a workflow file.
+It is the same trade [ADR 0014](0014-android-updates.md) accepted for the phone's nightly, and it is
+acceptable for the same reason and only that reason: this is not the channel anybody's real credentials are
+on. Rule 3 is untouched — the release channel still has no job, no token and no runner.
+
+**The version gets a floor, on this channel only.** MinVer answers `0.0.0-alpha.0.N` until the first `v*`
+tag and `vpk` refuses to pack anything below `0.0.1`, so the nightly job lifts the patch digit and keeps
+the prerelease height. It is applied through `MinVerVersionOverride`, which moves the assembly version too
+— packing a number the assemblies disagree with would put one version in the installer and another on the
+preferences screen, which is the screen somebody reads when asked which nightly they are running. Ordering
+across the floor holds: heights rise within a floored version, and the first real tag moves past all of
+them. The release script gets no floor and must not have one; there, `0.0.0` should be refused.
+
+### 11. Linux is a third desktop platform, and it is a package rather than an updater
+
+The Linux head existed and shipped as nothing: a checkout somebody built. `VelopackUpdateChannel` already
+named the two platforms it serves rather than negating the one it does not, on the reasoning that "a future
+AppImage should be an addition here rather than a thing that silently already half-happened". This is that
+addition, and it is not an AppImage.
+
+**The package is an RPM and nothing inside it replaces anything.** That is the decision, and everything
+else follows from it. Velopack does have a Linux path, so the shape of decisions 1 through 9 could have been
+repeated a third time — one more channel, one more index, one more client swapping its own files over. It
+should not be, because this platform already has the mechanism that whole apparatus is a substitute for.
+`dnf` is an updater the user already trusts, already audits and already controls, and it updates the
+runtime's own dependencies besides. Adding a second one that answers to this project instead would be
+building the capability decision 4 spends a page keeping unsettable, on the one platform where nobody had
+to.
+
+◆ **The first version of this rule said "there is no updater in it at all", and that went too far.** It
+conflated two things the other two platforms happen to do together: *noticing* that a newer build exists,
+and *installing* it. The second is what the argument above is about, and the first is not — a client that
+checks a fixed address and tells you what it found has taken no capability from anybody. Written as a flat
+refusal, the rule bought its safety by paying the cost in the paragraph below, and the cost was avoidable.
+
+**So the split is: this client checks and fetches, and PackageKit installs.** `LinuxUpdateChannel` reads
+the same rolling release the Windows nightly feed lives on, compares versions the way `rpm` compares them,
+downloads the package, verifies its digest, and hands the file to PackageKit — which raises polkit and asks
+the user, who may say no. `ApplyingEndsTheProcess` is false and nothing is torn down before the answer, which
+is the phone's shape rather than the desktop's, and is already what `MainWindowViewModel` does with that
+flag. Four things follow that are worth stating because each is a property the Velopack channels do not have
+to think about:
+
+- **The comparison is `rpmvercmp` and not SemVer.** Decision 7 made one version of record for the whole
+  repository, and this is the one place it is not the string that decides: `rpm` will accept or refuse the
+  install, so the client must compare the way `rpm` does or it can offer an update PackageKit then declines.
+  It is the same argument [ADR 0014](0014-android-updates.md) makes for comparing `versionCode`. The
+  manifest therefore carries both — the RPM version to compare and the SemVer to show a person — and the
+  installed side is read out of the rpm database rather than derived from the assembly, so the translation
+  `scripts/ci-linux-rpm.sh` performs exists in exactly one place.
+- **The digest is checked, and it is not a signature.** Velopack verifies what it downloads and Android
+  refuses a package whose signature does not match the installed one. An RPM fetched over HTTPS and handed
+  to a package manager has neither, so the manifest names a SHA-256 and the client refuses anything else.
+  That catches a truncated download and a caching proxy and says nothing whatever about provenance, because
+  the same job writes the package and the manifest. Provenance here is decision 4's constant address.
+- **`ONLY_TRUSTED` is cleared, deliberately.** The package is unsigned, so requiring PackageKit to trust it
+  would fail every update at the last step. Clearing the flag routes the request to polkit's
+  `package-install-untrusted` action, which asks for an administrator rather than proceeding quietly — the
+  user is told what they are agreeing to, which is the honest form of an unsigned package.
+- **Nothing restarts.** Replacing files under a running process is ordinary on Linux and the old inodes stay
+  mapped, so the build keeps running until somebody closes it. That is the same courtesy decision 1's
+  consequences argue for on Windows, arrived at for free.
+
+The honest cost that remains, recorded because it is the part that will be argued with: **discovery is still
+worse here than on either other desktop, for anybody who declines the in-app check.** There is no package
+repository to subscribe to, so `dnf upgrade` on its own will never bring a newer DodoSSH; the in-app check is
+what stands in for one, and a user who turns it off on PREFERENCES → UPDATES has no other signal. A signed
+repository is the fix, it is not built, and decision 11's last paragraph says what it would cost.
+
+**CI publishes it, and only to the nightly channel.** Rule 3 is untouched and applies with its full force:
+the token that writes a release is the capability to ship somebody a build, and it stays off runners. What
+changes is nothing, because rule 9 already established a channel CI may publish to and this is a second
+artefact on it. The release-shaped package is built on every `main` build and on every tag and uploaded
+nowhere, exactly as the Windows and macOS packages are, which proves the `linux-x64` restore graph and the
+packaging without putting an installer anybody has run in a place that invites passing it on.
+
+Note what CI's check is and is not. `vpk` verifies that the main executable actually calls
+`VelopackApp.Build().Run()` — a real assertion about the application, made for free by the packager. There
+is no equivalent here, so `scripts/ci-linux-rpm.sh` makes its own: it asks the finished package what it
+requires, what it provides, what files it holds and what modes they carry, and fails on each way a package
+can build cleanly and then be uninstallable or unsafe. Two of those checks exist because the first realistic
+payload tripped them — .NET's `libmscordbi.so` links `libmscordaccore.so`, which with bundled provides
+excluded produces a dependency no host can satisfy; and `cp -a` off the wrong filesystem produced a
+world-writable `/opt/DodoSSH`, which is a local privilege escalation against an application holding
+plaintext credentials.
+
+**The separations of rule 9 survive translation, and cost less here.** Two RPMs with different names owning
+different paths cannot replace one another and `dnf` says so, where on Windows the same property had to be
+built out of pack ids. The profile split is unchanged and still belongs to the build: `DodoChannel=nightly`
+is what `ClientPaths` reads, and the packaging must not answer that question a second time.
+
+**The device key store is settled separately and is not nothing.** The first version of this rule said this
+platform had none and that the passphrase was asked for at every unlock. It has one:
+[ADR 0007](0007-device-key-protection.md) decision 4 adds a TPM-sealed key released against a PIN the TPM
+itself checks and counts failures against. It is weaker than Hello and Touch ID in a way that ADR names
+precisely. Reaching the chip needs no group membership, because this package ships the udev rule that gives
+the seated user an ACL on `/dev/tpmrm0` — a decision that ADR overturned from its own first answer, and
+whose cost it states.
+
+**What this platform does not have is a signature on the package**, which is not decision 8's "unsigned for
+now": nothing on this platform
+refuses or warns about an unsigned RPM, so there is no dialog to buy a certificate to remove. What stands in
+for one is the same thing that stands in for it on the forge — TLS, and a project that publishes from one
+address. An RPM repository would change that, because repository metadata is signed and a repository is a
+thing users add once and then trust; that is the point at which a key becomes necessary rather than optional,
+and it is the reason building one is a decision and not a chore.
+
+**x86_64 only, for decision 10's reason and not a different one.** aarch64 is one argument to the publish
+and one to the script. What is missing is somebody to run Phase 19 on such a machine, and this project does
+not ship desktop builds nobody has run. The same sentence rules out claiming RHEL 9, which ships WebKitGTK's
+4.0 API line under a soname the package deliberately does not accept: `dnf` refusing with *nothing provides
+libwebkit2gtk-4.1.so.0* is a better answer than an install that draws a blank terminal.
+
 ## Consequences
 
 **The desktop gets what ADR 0011 had to refuse the phone.** Discovery is still manual — somebody has to be
@@ -281,6 +469,10 @@ ADR 0011 changes.
   genuinely the nicest onboarding available and will be proposed again.
 - **Conditioning updates on `MinClientVersion`.** See decision 5. It would create the attack rather than
   inherit it.
+- **A Velopack channel for Linux.** The mechanism exists — Velopack builds an AppImage — and repeating
+  decisions 1 through 9 a third time would have been the consistent-looking move. Declined in decision 11
+  because this is the one desktop platform that already has an updater the user controls, and a second one
+  answering to this project is the capability decision 4 exists to make unsettable.
 - **Azure Trusted Signing**, for now, on custody grounds rather than on mechanics. See decision 8.
 - **A hand-edited version property.** Considered seriously against MinVer, and it loses on the thing that
   matters here: the tag is already the version of record for the container image, and two places to write

docs/adr/0014-android-updates.md

@@ -116,6 +116,15 @@ is the entire reason there are two.
 other is an uninstall and a fresh enrolment. There is no migration and there will not be one — the two
 caches are encrypted under keys held by two package identities the platform keeps apart.
 
+◆ **The desktop has since taken this arrangement, and it had to build what Android is given.**
+[ADR 0013](0013-desktop-distribution-and-updates.md) decision 9 is this ADR applied to Windows: a nightly
+CI publishes from main, installing beside the release build rather than over it. The difference worth
+carrying back here is that every separation this ADR gets from the platform — different package identity,
+signature-checked updates, a per-app data directory — is on Windows a thing somebody had to choose and can
+therefore undo. The paragraph above about the attack surface applies there word for word, with one line
+removed: on the desktop the packages are not signed at all, so the feed is the only thing standing between
+a nightly install and an arbitrary build.
+
 **ADR 0011's "no automatic update" consequence is now wrong for the release channel and remains true for
 reach.** Discovery, MDM deployment and the sideloading permission are all unchanged. What changed is only
 that an installed copy can now learn a newer one exists.

docs/adr/0015-android-terminal-input.md

@@ -0,0 +1,246 @@
+# ADR 0015 — The Android terminal's input path leaves the WebView
+
+- Status: proposed — **Stage 1 built, unverified on a device**
+- Date: 2026-08-25
+- Builds on: [`docs/android-port.md`](../android-port.md) §9, [ADR 0011](0011-android-distribution.md)
+- Does not settle: whether the renderer stays xterm.js. That is [Stage 3](#stage-3--the-renderer-deliberately-unsettled), and it is deliberately left open.
+
+## Context
+
+The phone got the desktop's terminal, and the reasons were good ones. `TerminalDataPlane` already
+served the page and a binary protocol over a loopback WebSocket, the xterm.js bundles were embedded
+resources and platform-neutral, and `Avalonia.Controls.WebView` shipped a `net10.0-android36.0`
+target. `docs/android-port.md` called it *"better news than expected"*, and for rendering it was.
+
+**For typing it was not, and typing is the complaint.** Not scrollback, not colour, not throughput —
+the act of putting characters into a shell. That narrows the problem usefully, because the input path
+and the render path are separable, and only one of them is wrong.
+
+### Three things sit between a finger and a byte, and the middle one is the fault
+
+**1. Avalonia and the WebView fight over native focus.** `AvaloniaView.DispatchTouchEvent` ends every
+handled touch — DOWN and UP alike — with a `RequestFocus()` for Avalonia's own view. The WebView's
+input connection dies with its focus, so the keyboard swaps layouts and the inset churn can park it
+over the accessory row that was tapped. `Platform/TerminalFocus.cs` exists to hand focus back, posted
+to the next main-looper message so the platform has taken its turn first. It works. It is also a
+counter-move in a fight that only exists because the thing holding the input connection is not the
+thing the framework thinks is focused.
+
+**2. The keyboard is a prose keyboard, and no DOM attribute makes it stop.** This is the one that was
+worth checking rather than assuming, and the check inverted the obvious conclusion.
+
+The vendored xterm.js *already* sets every hint the web platform has on its helper textarea —
+`autocapitalize="off"`, `autocorrect="off"`, `spellcheck="false"`. They are set, and Gboard shows the
+suggestion strip and predictive text anyway. So this is not a missing attribute; it is a ceiling.
+
+Termux's `TerminalView.onCreateInputConnection` says what the ceiling is. It sets:
+
+```
+InputType.TYPE_TEXT_VARIATION_VISIBLE_PASSWORD | InputType.TYPE_TEXT_FLAG_NO_SUGGESTIONS
+```
+
+**`VISIBLE_PASSWORD` is the load-bearing half.** `NO_SUGGESTIONS` alone is advisory and most keyboards
+ignore it — a point made in Avalonia's own Android keyboard PR discussion, not only in Termux's. What
+actually turns off autocorrect, the suggestion strip and the predictive composing layer is the
+*variation*: a field the keyboard believes is a visible password gets no dictionary. That is an
+`EditorInfo.inputType` on a native `InputConnection`, and a page inside a WebView has no way to reach
+it — Chromium computes the `EditorInfo` from the focused DOM node, and no combination of attributes on
+a `<textarea>` produces the visible-password variation.
+
+**3. The accessory row is a workaround for a keyboard, not a keyboard.** Esc, Tab, Ctrl and the arrows
+are right and stay. But every tap on one costs the focus round-trip in (1), and none of them can be
+held or repeated.
+
+### The finding that decides this ADR
+
+Avalonia already emits Termux's exact `inputType`. Decompiled from the pinned
+`Avalonia.Android` 12.1.1, `AndroidInputMethod<TView>.OnCreateInputConnection`:
+
+```csharp
+if ((int)outAttrs.InputType == 1 && options.ShowSuggestions == false)
+{
+    outAttrs.InputType = (InputTypes)(outAttrs.InputType | 0x80080);
+}
+```
+
+`0x80080` is `TYPE_TEXT_VARIATION_VISIBLE_PASSWORD | TYPE_TEXT_FLAG_NO_SUGGESTIONS`. The guard reads:
+`ContentType` left at `Normal`, `Multiline` false, `AutoCapitalization` false, and
+`TextInputOptions.ShowSuggestions` explicitly `false` — a `bool?` whose default is `null`, so it must
+be set, not merely left alone. The same method also ORs `IME_FLAG_NO_EXTRACT_UI |
+IME_FLAG_NO_FULLSCREEN` into `imeOptions`, which is the other thing Termux sets and for the same
+reason: a terminal must never be replaced by a full-screen extract editor in landscape.
+
+So the keyboard behaviour that makes Termux feel like a terminal is **already available on the
+version this repository pins**, to any Avalonia control that asks for it. It is not available to the
+WebView, and never will be. That asymmetry is the whole decision.
+
+## Decision
+
+### Avalonia owns the keystrokes. The WebView renders and nothing else.
+
+A focusable Avalonia control — `TerminalInputSurface` — becomes the terminal's input target. It draws
+nothing. It sits in the same `Panel` as the renderer, takes focus when the terminal surface is shown,
+and keeps it for as long as that surface is up.
+
+- `TextInputOptions.ShowSuggestions = false`, with `ContentType`, `Multiline` and `AutoCapitalization`
+  left at the defaults the guard above requires. This is the entire fix for the suggestion strip.
+- `TextInputOptions.ReturnKeyType = Return`, which is the fix for Enter — see the correction below.
+- It implements `TextInputMethodClient` with an empty surrounding text and a zero-length selection.
+  A terminal has no document for an IME to edit; committed text is forwarded and nothing is retained.
+- Committed text and key events are encoded to bytes and written to the session through the same path
+  `TerminalScreen`'s accessory keys already use. No new transport, no change to `TerminalDataPlane`.
+- The native WebView is made non-focusable (`Focusable = false`, `FocusableInTouchMode = false`) once,
+  at attach.
+
+**`Platform/TerminalFocus.cs` is deleted by this, not amended.** The fight it referees ends when only
+one side is a focus candidate. Its remarks should move into the new control as the record of why the
+WebView is deliberately unfocusable, because that reasoning outlives the workaround.
+
+`Platform/SoftKeyboard.cs` stays. Dismissing a keyboard on the way to a sheet is still a platform
+call, and Avalonia still offers no way to close one.
+
+### Not: patching xterm.js's textarea
+
+Already tried by xterm.js itself, and already failing — see Context (2). The attributes are set. The
+platform ignores them. There is no version of this that reaches `VISIBLE_PASSWORD` from inside a
+WebView, so the effort buys nothing.
+
+### Not: a native Android `View`, Termux-shaped, today
+
+Termux's `TerminalView` + `TerminalRenderer` drawing to a `Canvas` is the right architecture for a
+phone-only terminal, and DodoSSH is not a phone-only terminal. Writing one now means two terminal
+implementations — this one and the desktop's — with two escape-sequence behaviours to keep honest
+against each other, for a defect that is entirely in the input path. That is a large bill for the
+wrong problem.
+
+It also does not become unavailable later. Stage 3 keeps it on the table.
+
+### Not: replacing xterm.js now
+
+Rendering is not the complaint. xterm.js with the WebGL addon draws a correct, fast grid, and
+`terminal.js` already handles renderer death, reattach and refit — hard-won behaviour recorded in
+`docs/android-port.md` that a rewrite would have to re-earn. Replacing a working renderer to fix a
+keyboard is the wrong trade, and doing it in the same change as the input move would make neither
+half attributable when the result is judged on a device.
+
+## Staging
+
+### Stage 1 — the input surface · **built**
+
+The decision above, and nothing else. This is the change that answers the complaint, and it should be
+judged on a device on its own before anything else is built on it.
+
+What landed:
+
+| | |
+| --- | --- |
+| `Views/TerminalInputSurface.cs` | The 1×1 focusable control, `ShowSuggestions = false`, the IME client that retains nothing, and the key→escape-sequence table |
+| `Views/TerminalInputEventArgs.cs` | One keystroke's bytes, on the way to `TerminalScreen.SendAsync` |
+| `Platform/RendererFocus.cs` | Makes the `WebView` non-focusable and reports taps on it |
+| `Platform/TerminalFocus.cs` | **Deleted.** The fight it refereed has one contestant now |
+| `Views/TerminalScreen.axaml{,.cs}` | Hosts the surface, raises the keyboard on arrival, drops the two posted focus counter-moves from the accessory row |
+
+**One thing gained that was not asked for, and it was a real defect:** the Ctrl latch now reaches the
+software keyboard. It never did before — everything typed went straight into the page and never passed
+`SendAsync`, which is where the latch lives — so `ctrl` then `c` was inert unless the second key also came
+from the accessory row. Check 11.10d covers it.
+
+**Verify before building anything on this:** checks 11.10a–11.10e in `docs/manual-checks.md`, and 11.10b
+first of all, because it is the falsification test at the end of this document.
+
+> **⚠️ Corrected on the first device run: every Enter closed the keyboard.** The premise held — the
+> suggestion strip, autocorrect and predictive text were all gone — but the field had been left with
+> Avalonia's default `ReturnKeyType`, and the same `OnCreateInputConnection` maps that, on a field that is
+> not `Multiline`, to `IME_ACTION_DONE`:
+>
+> ```csharp
+> _ => options.Multiline ? 1073741824 : 6,   // 6 is IME_ACTION_DONE
+> ```
+>
+> Done means done: the keyboard performs the action and dismisses itself. On a text box that is right, and
+> on a terminal it fires on every command. `TextInputReturnKeyType.Return` — value 1, so
+> `returnKeyType - 1 => 0 => 1073741824` — is `IME_ACTION_NONE`, which leaves Enter a plain newline with no
+> action for the IME to perform and no reason for it to hide.
+>
+> **The trap worth recording, because it is the same trap 11.10b already warned about from the other side:**
+> `Multiline = true` also reaches `IME_ACTION_NONE`, through the default arm, and is the wrong way there. It
+> ORs `TYPE_TEXT_FLAG_MULTI_LINE` into `InputType`, which moves it off `1` and silently fails the
+> `ShowSuggestions` guard. Taking the obvious route to a working Enter would have handed the suggestion
+> strip back and looked like the whole ADR had failed. `ReturnKeyType` touches `imeOptions` only.
+>
+> Enter's *byte* was never in doubt and is not what changed — `IME_ACTION_DONE` was already synthesising
+> `KEYCODE_ENTER` through `AvaloniaInputConnection.PerformEditorAction`, so commands ran; the keyboard just
+> left afterwards. `OnTextInput` now also normalises line endings to CR, for the case where a keyboard with
+> no editor action commits `"\n"` as ordinary text instead of sending a key event, and so that a pasted
+> `"\r\n"` does not run its line twice.
+
+### Stage 2 — the keys around it
+
+Only once Stage 1 is proven, because each of these is cheap *given* Avalonia owning input and awkward
+without it:
+
+- **Hold-to-repeat** on the arrows and Backspace. Currently impossible without a focus round-trip per
+  tap.
+- **Ctrl held rather than latched**, with latching kept as the fallback. `TerminalScreen`'s remark is
+  right that latching is what phone SSH clients do; held becomes possible when one control sees both
+  the modifier press and the following keystroke.
+- **Pinch and volume keys for font size**, replacing the pinned `A−`/`A+` buttons. Both are what
+  JuiceSSH does and what its users name first; the buttons stay as the discoverable path.
+
+Volume-key capture is an `onKeyDown` on the activity and is deliberately opt-in — a phone that cannot
+change its volume with a shell open is a worse phone.
+
+### Stage 3 — the renderer, deliberately unsettled
+
+If, after Stages 1 and 2, the terminal still reads as a web page — touch selection, scrollback
+survival across renderer death, scroll feel — then the renderer is the next question, and the
+candidates are known:
+
+| Option | What it is | Cost |
+| --- | --- | --- |
+| [XTerm.NET](https://github.com/tomlm/XTerm.NET) | MIT, headless, UI-agnostic; VT100/ANSI, dual buffers, 256/true colour, wide chars | Evaluate emulation fidelity against xterm.js |
+| [VtNetCore](https://github.com/darrenstarr/VtNetCore) | MIT, .NET Standard 2.0, older but proven | Maintenance status |
+| Port xterm.js's core | Exact behavioural parity with the desktop | Largest, and a permanent fork |
+
+Rendered by a custom Avalonia `Control` via `DrawGlyphRun`, which is Termux's architecture expressed
+in Avalonia and would work on the desktop too — eventually letting the WebView2 dependency go, and
+with it the `targetSdk` 36 floor `Avalonia.Controls.WebView` currently forces.
+
+**Stage 1 is not wasted if Stage 3 happens.** The input surface is exactly the layer such a control
+would need, unchanged. That is the reason for splitting them.
+
+## Consequences
+
+**What gets better, and it is one thing done properly.** No suggestion strip, no autocorrect, no
+predictive composing layer, no full-screen extract editor, and no focus flap on an accessory key.
+That is the difference between typing at a shell and fighting a phone.
+
+**CJK and other composing input is degraded, and this is a real cost.** `VISIBLE_PASSWORD` suppresses
+composition, which is precisely what makes it work for a terminal and precisely what a Chinese,
+Japanese or Korean user needs. Termux has the same limitation and lives with it. If it ever needs
+answering, the answer is a per-session toggle that drops `ShowSuggestions = false` and accepts the
+prose keyboard back — not a middle setting, because there isn't one.
+
+**Text selection and copy are unchanged**, still xterm.js's own DOM handling, still without touch
+selection handles — [xterm.js #5377](https://github.com/xtermjs/xterm.js/issues/5377) is open and
+unassigned and this ADR does not fix it. Stage 3 territory.
+
+**Scrollback still dies with the renderer process.** Unchanged, still honest about it via the replay
+banner, still Stage 3 territory.
+
+**Nothing here is verifiable headless**, for the reason `docs/manual-checks.md` §11 already gives.
+Stage 1 needs new manual checks — at minimum: the suggestion strip is absent with a shell open; an
+accessory key does not disturb the keyboard (§11.10a, which should now pass trivially rather than by
+posted counter-move); landscape does not produce a full-screen extract editor; and a hardware
+keyboard still reaches the shell.
+
+**The hardware-keyboard path moves.** It currently arrives through the WebView; after this it arrives
+through Avalonia. Behaviour should be identical and probably will not be on the first try.
+
+## What would falsify this
+
+If, on a device, `ShowSuggestions = false` does not suppress the suggestion strip — because the guard
+was not satisfied, or the keyboard ignores the variation too — then Stage 1's premise is wrong and
+the honest next step is Stage 3 with a native `InputConnection`, not more Avalonia. **Check this
+first, on the smallest possible spike, before building `TerminalInputSurface` properly.** It is one
+attached property on a focusable control and a look at the keyboard.

docs/android-port.md

@@ -30,7 +30,10 @@ verified is that it compiles, links, packages, and carries the right natives.
 transfers protected by a **foreground service**. File transfer is not in the first scope; when it arrives it
 is **one remote pane** with Android's document picker for moving files in and out. *It has since arrived,
 both ways:* the pane, the queue, `ACTION_OPEN_DOCUMENT` going in and `ACTION_CREATE_DOCUMENT` coming out,
-with the foreground service now counting transfers as well as shells.
+with the foreground service now counting transfers as well as shells — and, since, an idle-but-connected
+Files session as well, which a transfer count alone was blind to. *Corrected the same round:* the service's
+other half — a shell's own opening — had never been wired to anything at all, so a shell survived only for
+as long as the app stayed foreground; see [Sessions survive backgrounding](#sessions-survive-backgrounding-via-a-foreground-service).
 
 **What was actually checked**, so the rest can be read with the right amount of trust:
 
@@ -235,6 +238,43 @@ The parts that are definitely different are the on-screen keyboard, and the fact
 needs Ctrl, Esc, Tab and arrows that the software keyboard does not offer — every Android SSH client ships an
 accessory key row for this. That is UI work, not porting.
 
+> **⚠️ Corrected by the build. The data plane assumed a renderer that attaches once and lives forever, and
+> that assumption is WebView2's truth, not Android's.** Desktop's WebView2 process starts with the window and
+> dies with it; `TerminalDataPlane` was written to that reality — one socket, attached once,
+> `Interlocked.Exchange`-guarded against a second attach ever happening at all. On a phone the WebView's own
+> renderer process is a separate thing from the app process the foreground service above is keeping alive,
+> and Android kills *that* independently — under memory pressure, or simply for being backgrounded — with no
+> foreground service able to save it. The page then reloads with a fresh socket, and three things broke on
+> that reload before this was found: the second attach was refused outright (`409 Conflict`), because a
+> second valid upgrade could only mean a bug or a hostile second process, never our own page coming back; a
+> send into the dead first socket threw, and that exception unwound `TerminalSessionPump`'s flush loop,
+> freezing the still-live shell behind it — `LiveSessionCount` kept counting a session nothing would ever
+> drain again; and every byte sent while no page was attached had already spent flow-control credit that no
+> acknowledgement could ever return, so a session outliving 256 KiB of output into a dead page stalled for
+> good regardless of the other two. Waiting for the old socket to notice it was dead and close on its own
+> was never going to be enough either — a killed renderer sends no TCP FIN, so the old receive loop could sit
+> unaware for the whole 30-second keepalive.
+>
+> Fixed as a takeover rather than a guard: a second valid upgrade — origin, token and subprotocol all
+> checked exactly as before — now displaces whatever socket was attached instead of being refused, since
+> only this app's own page ever knows the token, so a second valid attach *is* that page, back again.
+> `TerminalDataPlane.SendAsync` no longer lets a dead-socket send escape as a fault; it reads as "nobody
+> listening," same as no socket being attached at all. `TerminalWorkspace` resets each live session's credit
+> window on every attach and resends its `SessionOpened` frame, flagged as a replay, so the fresh page
+> rebuilds the pane and the pump stops waiting on an acknowledgement that was never coming. And
+> `terminal.js`'s socket now retries itself, forever, with backoff, instead of reporting the connection
+> failed and stopping — the page dies with the app anyway, so there is no case where retrying is the wrong
+> call. What is **not** recovered, and says so rather than pretending otherwise: scrollback across a page
+> reload. It lived in the page's own DOM, and a reloaded page is a new DOM. The replay banner — *"the view
+> reconnected; earlier output stayed on the host"* — is that honesty put where the person looking at the
+> terminal will actually read it, not buried in a log.
+
+> **⚠️ Under review, on the input half only.** The renderer's story above stands; what has not held up in
+> use is *typing*. A software keyboard over a WebView cannot be told to stop autocorrecting — the DOM
+> attributes xterm.js already sets are advisory and Android ignores them — and the accessory row pays a
+> focus round-trip per tap. [ADR 0015](adr/0015-android-terminal-input.md) proposes moving the input path to
+> Avalonia and leaving the renderer exactly as this section describes it.
+
 ---
 
 ## Decisions taken
@@ -280,13 +320,50 @@ What is desktop-only is the *left* pane — `LocalDirectory`, the drive list, th
 
 ### Sessions survive backgrounding, via a foreground service
 
-A persistent notification for as long as a shell or a transfer is live.
+A persistent notification for as long as a shell, a transfer, or an idle-but-connected Files session is
+live.
 
 It costs the user a notification and some battery. It buys the behaviour the desktop client already promises
 and documents — that a shell outlives a vault lock, and that a transfer finishes — and the alternative was
 to make `TerminalWorkspace`'s guarantee desktop-only, which is a worse thing to have to write down than a
 notification is to look at.
 
+**Three corrections found after the first cut shipped, all in the wiring rather than the design:**
+
+- **A shell opening never started the service.** `SessionKeepAlive` heard `TerminalWorkspace.SessionEnded`
+  and refreshed on that, but nothing announced the opposite event — so a user who opened a shell and
+  backgrounded the app immediately had no foreground service at all, and Android was free to kill the
+  process holding it. `MainWindowViewModel.TerminalSessionOpened` is now wired the same way in
+  `App.axaml.cs`'s `ComposeKeepAlive`.
+- **A connected-but-idle Files session counted as nothing.** A host open on the Files screen with no
+  transfer moving is a live SFTP connection a dying process would sever, and the old two-argument
+  `Reconcile(liveSessions, activeTransfers)` had no way to hear about it. `TransfersViewModel.HasLiveFileSession`
+  — `IsConnected` with a real `ConnectedCipher`, which a bucket never has — is the third fact `Reconcile` now
+  takes.
+- **Refreshing the notification restarted the service, which throws when backgrounded.** `Reconcile` called
+  `StartForegroundService` on every refresh, including the common case of a service that was already
+  running. On API 31+ that throws `ForegroundServiceStartNotAllowedException` the instant the app is
+  backgrounded — a transfer finishing in the pocket, one of two shells dying — which crashed the process and
+  took every session with it. `SessionForegroundService` now tracks whether it is already running and, when
+  it is, posts the updated notification through `NotificationManager.Notify` instead of asking Android to
+  start anything.
+
+**The notification permission is requested, not just declared.** API 33+ requires `POST_NOTIFICATIONS` at
+runtime or the receipt is silently invisible — the service still runs, but nothing on screen says so.
+`SessionForegroundService.Reconcile` asks for it the first time in this process there is actually something
+to show, at most once, with no result read back: a refusal costs the notification and nothing else, which is
+what the manifest's own comment on the permission says.
+
+**And a fourth correction, found by the notification refusing to come down.** "1 shell connected" outlived
+the shell, both ways a shell can close. A shell exiting on its own announced `SessionEnded` from inside its
+run's own finally block — where the run task is by definition not yet complete, so the
+`LiveSessionCount` the keep-alive reads from that event still counted the session that had just ended, and
+nothing fired afterwards to correct it. A tab closed by hand announced nothing at all, by a recorded
+decision that assumed every subscriber was the closer. Both reversed in `TerminalWorkspace`: the end is now
+announced from a continuation after the run has actually completed, and `CloseSessionAsync` announces too,
+after its own drain — the event's remark carries the reversal, and `SessionEnded`'s subscribers were all
+already "reconcile to reality" handlers for which a second announcement is harmless.
+
 ### Phone first
 
 About 360dp wide. The tablet route was cheaper — a landscape tablet is close to the existing 880×560 minimum
@@ -523,7 +600,12 @@ go at 360dp:
    stopping it from a count rather than a lifecycle. `TerminalWorkspace.LiveSessionCount` is the source of
    truth deliberately: it already knows that a session whose shell exited is not live, which a counter
    incremented on open would not, and a phone showing "1 shell connected" over nothing would be exactly the
-   dishonesty the unlock screen's count exists to prevent.
+   dishonesty the unlock screen's count exists to prevent. *Corrected since:* the opened half of a shell's
+   lifecycle was never wired in, so the service could never come up for a shell at all; an idle-but-connected
+   Files session now counts as a third live fact rather than nothing; a refresh while backgrounded updates
+   the notification in place instead of restarting the service, which the API throws on; and
+   `POST_NOTIFICATIONS` is now actually requested rather than merely declared. See
+   [Sessions survive backgrounding](#sessions-survive-backgrounding-via-a-foreground-service) for all four.
 7. ~~**The interface**, phone-first.~~ **Done for the decided scope** — all seven screens of the design,
    plus the two states the design does not draw because it starts at an enrolled phone (naming a server, and
    choosing a passphrase).
@@ -568,9 +650,16 @@ go at 360dp:
    believing they typed an `l`.
 
    **The surface has since taken the whole screen.** `PhoneShell` collapses the header, the session strip
-   and the bottom bar while a terminal is showing — one binding on `IsShowingPages` each — and the screen
-   draws a 35-pixel bar in their place: back, the session pills, and a `+` raising a sheet with the three
-   connections there are. That sheet is the head's first control that could be drawn over the renderer, so
+   and the bottom bar while a shell is showing, and the screen draws a 35-pixel bar in their place: back,
+   the session pills, and a `+` raising a sheet with the three connections there are.
+
+   *A shell rather than the surface, and the two parted company once that surface gained a connect page.*
+   With nothing running, Connections is a box, a CONNECT button and the machines connected to before — a
+   page in everything but which enum it is in — so `RefreshChrome` keeps the bar (and, wide, the rail) under
+   it and the Connections entry lights for the first time. It is the one screen reachable by closing your
+   last tab, and collapsing the nav there left the system back gesture as the only route to Hosts or
+   Settings. The header is not part of that: the surface draws its own bar, and a vault header above it is
+   the second row of chrome this head exists to avoid. That sheet is the head's first control that could be drawn over the renderer, so
    it collapses it rather than covering it, exactly as the desktop's palette does; whether Android's
    WebView actually composites above Avalonia content is still the unverified question recorded below, and
    collapsing is correct under either answer.
@@ -666,6 +755,12 @@ Recorded so they are choices rather than accidents. Any of them is cheap to revi
   local-and-non-roaming requirement more cleanly than any desktop platform does.
 - **The device name** on log entries comes from the head, not `Environment.MachineName`, which returns
   something like `localhost` on Android and would make every entry from a phone indistinguishable.
+- **The Hetzner Cloud importer is not ported, and it is the one gap here that is not a platform limit.**
+  Its view model is in `DodoSSH.Client.Shell` like every other and its reader is an `HttpClient` call, so
+  both halves already run on a phone — unlike the `~/.ssh/config` importer beside it, which has no file to
+  read under scoped storage and never will. What is missing is a layout: the desktop draws a five-column
+  table of servers, and a phone version of that is a different screen rather than the same one narrower.
+  Said on the Preferences screen rather than left to be discovered.
 - **`NativeKeyboardFocus` is not ported.** It exists for a documented Win32 asymmetry — focus crosses into
   WebView2 but does not come back — and Android's focus model is different enough that the problem should be
   confirmed to exist before anything is written to solve it.

docs/design-import-gaps.md

@@ -69,7 +69,7 @@ the chrome, hosts and terminals, file transfer, the vault, teams, and preference
 > labelled sidebar and the chrome went from 72 tall to 86, so `880x560` became `1016x574` — leaving every
 > screen the same `826x464` it was designed against. Four of the tables stop fitting at 690 wide, so
 > widening the sidebar without widening the window would have broken them where the layout suite was not
-> looking.
+> looking. **v5b grows it again**, to `1081x583` — see that section, below.
 >
 > **Buckets became a destination** rather than a toggle inside the files screen, matching the phone: the
 > `HOST` / `BUCKET` pair is gone and `ShellScreen.Buckets` draws the same `TransfersScreen` with the other
@@ -102,6 +102,14 @@ the chrome, hosts and terminals, file transfer, the vault, teams, and preference
 > each get the full window width instead of `826`. See `MainWindowViewModel.IsVaultsTab` for why the tab is
 > a page test rather than a fourth `ShellSurface`.
 >
+> **v5b reverses the paragraph above, and says so rather than leaving it to be found out of date.** The tab
+> strip this paragraph describes is retired — see the v5b section, below — and with it the rule that made the
+> rail conditional at all: the switcher it drew moved onto the nav rail's own head as three segments, the
+> rail stopped collapsing for any of them, and neither SFTP, S3 nor a terminal gets the window's full width
+> any longer. `IsVaultsTab` outlived the strip essentially unchanged — it is still a page test rather than a
+> read of a `ShellSurface`, now over which of the rail's own destinations is showing rather than which of
+> three tabs was.
+>
 > **The hosts screen became a grid of cards** — groups above, hosts below — and the 268-pixel host sidebar
 > went with it. That column was choosing among forty machines *and* editing one of them at two-thirds
 > width; the grid took the first job at full width and a 304-pixel right-hand drawer took the second. The
@@ -160,11 +168,214 @@ the chrome, hosts and terminals, file transfer, the vault, teams, and preference
 > | **Share this host** | Omitted, as the vault screen's `SHARED WITH · 6` is. A grant is per *vault* and per-item sharing is M5; a button here would imply this one host could be handed over, which is the thing the architecture does not do. |
 > | **Add Telnet**, and **Serial** in the toolbar | Omitted. `ISshConnection` is the only transport there is. This is also why the card subtitle's `ssh` is a constant today rather than a reading — it is stated in `HostRowViewModel.Summary`, which is the one place in this interface where a constant is printed on purpose. |
 > | **+ SSH ID, Certificate, FIDO2** | Omitted. `IDENTITIES` and `CERTIFICATES` have been on this document's list since the first import — neither is even a reserved `SyncEntityType` — and there is no security-key path anywhere in the SSH layer. One control offering three item types that do not exist. |
-> | The **Backspace / Default** row | Omitted. It is a terminal setting, and the client has no preferences store and no frame to carry one to the renderer — see the Preferences section. It would be a control whose value could not survive the window closing. |
+> | The **Backspace / Default** row | Omitted, and it is the one row on this list that could now be built cheaply — a setting, an opcode and a control, the way the text size was. What it lacks is a reason: which byte backspace sends is a fact about the remote's `stty`, so a client-side switch fixes a mismatch by hiding it. See the Preferences section. |
 > | The **chevron beside the vault name** | The name alone, and the move behind the pane's ⋯ menu instead. A host *can* now be moved between vaults, so the gap is no longer that there is nothing to offer — it is that a chevron on a subtitle implies an edit, and this is not one: the two vaults are encrypted under different keys, so it is a re-seal into one and a tombstone in the other, the host takes a new id, and its group and tags stay behind. A control that implied "just change this field" would be describing something else. Where a *new* host goes is still asked in the host editor, as a picker beside the name. A group moves too, from its card's right-click menu, and takes its nested groups and every host filed under them; keys, passwords and buckets take theirs from the keychain screen's standing picker and cannot be moved yet. |
 > | **Show more ⌄** | Not drawn as a disclosure. What it would hide — notes, the relay switch, forgetting the host key — is in the editor, one press away, and a second fold inside a pane that already scrolls is a second place for a field to be missing from. |
 > | **Port Forwarding** in the sidebar | Nothing, for the third time in this document. |
 > | The host grid's toolbar avatar, share and tag-filter controls | Omitted, as in v3 and for the same reasons. |
+>
+> ## The desktop's v5 — hosts, restyled, and two reversals
+>
+> A fifth pass, and the first that undoes as much of its own predecessors as it adds: a re-theme both heads
+> share, pinned folders as a full feature, a last-connected time read off the connection log, and two pieces
+> of earlier passes put back on the user's own instruction rather than kept.
+>
+> **The palette went purple, and three fonts are embedded that were only ever named before.** The shared
+> scheme moves from v3's near-black-and-blue to a purple accent (`#5D42DE`, with a gradient and a glow of its
+> own) over the same near-black canvas, and both heads recolour together — `Palette.axaml` is shared, and a
+> colour is not allowed to mean something different on the two of them. What changed on the desktop alone is
+> the type: Montserrat, JetBrains Mono and Material Icons are embedded as `AvaloniaResource` in the Shell
+> project rather than requested by name and left to whatever the machine happens to have, the way `WithInterFont`
+> alone used to leave Inter registered but unused. Montserrat is the default sans now, with Inter kept as its
+> fallback rather than the whole answer; `MonoFont` gains JetBrains Mono at the front of its own list, ahead of
+> the system-mono stack that is still behind it for a glyph JetBrains Mono does not cover. **Android recoloured
+> with the shared palette and kept its own default sans when this paragraph was first written; it does not any
+> more** — see "the phone catches up", directly below.
+>
+> **The phone catches up.** A later reversal than the one above, on the same reasoning: a face is not a
+> colour, and the user decided the phone should share both rather than only the second. `WithInterFont` gains
+> the same `FontManagerOptions` block `Program.cs` sets, off the same embedded Montserrat, no new asset
+> required. `Theme/Phone.axaml`'s six-rung v2 ladder — 4/9/10/11/12/14 — collapses to the desktop's three:
+> chip or tag 6, button or field 10, card or output 12; `Button.fab`'s 28 and every sheet's 22 survive
+> unchanged, being geometry rather than ladder rungs. `Button.primary` and `Button.fab` take `AccentGradient`
+> and `AccentGlow` in place of a flat fill, the same swap `Button.accent` made on the desktop, `BoxShadow`
+> cleared the same way on disabled. And the phone's own chrome — the vault header, the bottom bar, the shells
+> strip, `HostActionBar`, the editor headers, the terminal's collapsed bar and every screen's raised selection
+> bar — moves from `Chrome`/`Sidebar` to `DeepChrome`, joining the desktop's own v5b titlebar-and-rail move;
+> `PhoneRail`'s hover takes `Track`, which `Palette.axaml` already names for that row. Borders stay keyed.
+>
+> **Flat sections replace the grid of group cards and the breadcrumb trail — reversing v3, on the user's own
+> approval rather than a defect found in it.** v3's grid held one level of the group tree at a time, opened by
+> a double-click, with a trail above it saying where you were; that model is retired outright rather than
+> folded into anything; every group is now a heading on one flat board, in label order, exactly the flattening
+> the phone's `SidebarRows` has always drawn — a headerless board for a groupless keychain, "No group" first
+> where there is one, a chevron per heading and one Collapse All. Nothing here was found broken; the grid did
+> what v3 asked for. The user asked for the phone's shape on the desktop too, and decision 2 in
+> `hosts-v5-design-spec.md` is that request recorded.
+>
+> **Tag chips and a mono address line are back on the card — reversing v4, also by approval and not by
+> defect.** v4 had replaced both with a two-line `ssh, root, pci, eu-west-1` subtitle and moved the address to
+> the tooltip alone, on the reasoning that a card read while scanning forty machines wants the kind of machine
+> more than its address. v5's own mock draws chips and an address on every card, and the user chose the mock's
+> arrangement over v4's when the two were put side by side — so the card now carries both: the full address is
+> still on the tooltip, as v4 left it, and a truncated mono copy sits under the name as well.
+>
+> **Pinned folders are a full feature, not a field with nowhere to be edited.** `HostSecret.PinnedPaths` is an
+> ordered, deduplicated list merged per path the way `TagIds` is merged per tag, so two machines pinning
+> different folders on one host both keep theirs. The drawer's QUICK ACCESS section edits it — a row per pin
+> with a close box, an add field and button, staged on `VaultViewModel.EditorPinnedPaths` the way the tag
+> picker stages `editorTagIds` — the card shows a pin-count badge once there is at least one, and the detail
+> pane repeats the list read-only in its own fieldrow idiom. Clicking a pin above a connected terminal opens
+> the SFTP tab with the remote pane navigated straight to that path, through the same connect path "Browse
+> files" already used — a second authenticated connection, exactly as every SFTP session here is.
+>
+> **Last connected is read off the synced connection log, and it is deliberately narrow about when.**
+> `max(StartedAt)` per `HostId` is computed on the hosts screen's own activation and again when a session ends,
+> then restrung into words on a one-minute tick for as long as the screen stays visible — never on the sync
+> loop, which is the same rule `LogsViewModel` states for the log itself: nobody is watching their own
+> connection from an hour ago update on a screen they are not looking at. A connected host prints nothing here;
+> the green dot already answers "right now", and printing an age beside it would be answering a question about
+> a session that is not the one still running. **It is scoped to the active vault alone**, the same limitation
+> `LogsViewModel` and `MainWindowViewModel.RecentConnections` already carry — a host filed in a second vault
+> this session can read gets no ago-text until something on this screen reads more than one vault's connection
+> history, which nothing does yet. And a session ending can, briefly, still show the previous entry: the
+> recorder's own write is queued onto a background task rather than made inline, so a re-read landing before
+> that write drains reads the log as it stood before the session it is reporting on closed. It self-heals on
+> the next activation or the next session end, and closing the gap outright was judged not worth blocking a
+> screen transition on the recorder's queue.
+>
+> **Quick connect grew to match, and quietly answers a different question than the mock's own row does.** The
+> card widened from 520 to 640, gained the mock's `>_` prompt styling and a live dot per row the way the hosts
+> board's own two-state dot works. What it does not draw is the mock's SSH/SFTP kind column — every palette
+> connection here is SSH, so a constant printed as though it varied would be exactly the kind of decoration
+> this document has refused before (the terminal's session footer, the snippet's kind badge); the auth word
+> (credential/key/password) answers a question that is actually different per row instead.
+>
+> | v5 element | What ships instead |
+> | --- | --- |
+> | The status dot's third, amber state | ◆ **Still two states, on the same reasoning v3 and v4 both gave and v5 was asked to reconsider.** Green means a terminal is open on that host, grey means one is not; nothing here pings a machine, so the amber "reachable but not connected" state has no fact behind it to draw. |
+> | The relay checkbox's mock copy | Restyled into the mock's own nested-card shape (radius 12, a title beside the box rather than under it). The "(not wired up yet)" it carried while this client could not dial the relay came out when it could; the sentence under it now says what ticking it does — the server dials the host, so the server has to be able to reach it, and the address is stored in plain text. |
+> | The host card as a link straight to a terminal | Click still selects, double-click still connects, and the pencil still opens the pane — the mock's card-as-link is not adopted, because multi-select (Ctrl, Shift, the marquee band) depends on a plain click meaning "choose this one" rather than "go". |
+> | Quick connect's SSH/SFTP kind column | The auth word — credential, key, or password — see above. |
+> | A collapsed section staying collapsed after a restart | In memory only, for the running session. `settings.json` holds two scalars by decision — the terminal's text size and whether this machine checks for updates on its own — and collapse state is not judged worth a third. |
+> | QUICK ACCESS's editor, on the phone | ◆ **Shipped, over the same shared data the deferral above described.** The phone's host editor draws its own QUICK ACCESS section on the same staged `VaultViewModel.EditorPinnedPaths` the desktop's drawer binds — a row per pin, an add field and button, and a remove target sized to this head's 44dp touch floor rather than the desktop's 22-pixel close box. `AddEditorPinCommand`'s refusals surface through a `Status` line the editor page draws for itself, since that page covers the whole screen and the list behind it draws its own `Status` off-screen for as long as it is open. The pins themselves reach a second surface this head has that the desktop does not need: `TransfersViewModel.ConnectedPinnedPaths` carries them as chips on the Files screen while connected, and tapping one runs `GoRemoteCommand` — the same command the breadcrumb trail already used to navigate. Two deviations from the desktop, both named where they land: the chip row is a snapshot taken at connect rather than a live follow of the vault, so a pin edited mid-session shows up on the next connect rather than this one; and the editor's own hint sentence says the pins appear on the Files screen, not above a terminal — this head has no terminal strip for them to sit above, the same honesty the row below already states for the desktop's own hint. |
+> | Collapse All beside every section's own collapse chevron | Bound on every heading's view model and shown on only the first — `SidebarGroupHeader.IsFirstBoardSection` is what a virtualised list of sections uses in place of a control of the board's own that would otherwise have to sit above all of them. |
+> | The QUICK ACCESS hint's claim that pins live in a sidebar | "Pinned folders appear above the terminal for this host." — no sidebar exists on this screen for the sentence to point at, so the shipped hint says where they actually draw. |
+> | The mock's "Saving to **DodoTech ▾** vault" subtitle, with a picker's chevron inside a sentence | The pre-existing `DrawerSubtitle` wording — the vault's name alone, unchanged by this pass. A chevron inside running text implies the text itself is the control, which it is not: the vault picker is its own element, shown only while creating and only above one writable vault, as it always has been. |
+> | The design's two deeper text steps, `#6D6F84` and `#5D5F74` | Not added as `Palette.axaml` keys. `TextGhost`, already repicked to `#7C7F98` for v3, is reused everywhere the design reaches for either — a resource nothing yet distinguishes from `TextGhost` is not free to carry, on the same reasoning the border ramp's own remarks give. |
+>
+> ## The desktop's v5b — chrome, session shell and SFTP restyled
+>
+> A sixth pass, and the first aimed at the chrome and the two screens somebody spends the most time inside
+> rather than at any one screen on its own. **What shipped:** the titlebar and nav rail redrawn against
+> `TitleBar.dc.html` and `NavRail.dc.html` — a segmented SSH/SFTP/S3 switcher at the rail's own head, a
+> mode-dependent first row beneath it, and Vaults and Preferences moved off the rail's list and into a
+> popover under the user chip at its foot. The terminal and the SFTP surface both gained a session shell — an
+> in-screen tab row, a host header, a bottom status bar and a 300-pixel QUICK ACCESS/SNIPS sidebar —
+> replacing the window-wide tab strip v3 built and the pin-chip strip that used to sit above the terminal
+> alone. The SFTP screen's own two panes and its TRANSFERS strip were restyled against `SFTP.dc.html` — row
+> grids, quiet Material glyphs, slim progress bars — and three screens that had never had a `.dc.html` of
+> their own got one each: Keychain, Snippets (titled Snips on screen now, matching the rail) and Logs.
+>
+> **The nav rail stopped being conditional furniture, and that is the one structural change underneath the
+> paint.** v3 drew it only under the Vaults tab, so SFTP, S3 and an open terminal each got the window's full
+> width; v5b's own fidelity pass puts the rail up beside all three, because the design draws it that way and
+> because the segmented switcher now living at its head is only worth having if it stays reachable from
+> everywhere. `MainWindowViewModel.IsVaultsTab` still exists and still answers the question it always did —
+> which of the rail's own six destinations is on screen, as opposed to SFTP or S3 — but nothing about it gates
+> the rail's own visibility any more. The window's minimum grew again to hold the difference: `1016x574`
+> became `1081x583`, nine pixels for the titlebar's 44-to-53 and sixty-five for the rail's 190-to-255, added
+> straight onto the minimum rather than absorbed by shrinking a screen — see `MainWindow.axaml`'s own remark
+> and `LayoutHarness.MinimumWidth`/`MinimumHeight`. `ScreenWidth` held at 826 because the rail is the only
+> thing beside a page that grew; `ScreenHeight` grew by the 42 pixels the retired tab strip used to cost every
+> screen, because nothing replaced that strip as chrome every screen pays for — the terminal and SFTP
+> surfaces pay for their own tab row out of their own session-shell budget instead. See
+> `LayoutHarness.ScreenHeight` and `SessionScreenHeight`.
+>
+> | v5b element | What ships instead |
+> | --- | --- |
+> | The host header's OS label (`Ubuntu 24.04 LTS`) and latency reading (`12 ms`) | Neither. The client does not know the remote's OS and nothing measures round-trip time — the same two absences the terminal pane header already recorded before this pass, carried into the new header rather than reopened. |
+> | A **Port forward** button on the host header | Omitted. The feature does not exist; see the FORWARDING rows earlier in this document. |
+> | The tab and status dots' third, amber state | ◆ **Still two states**, for the reason the hosts screen's own dot has stayed two states since v3: green means a shell is open (or a session is connected), grey means it is not, and nothing here pings a host to justify a third colour meaning "reachable but not connected". |
+> | The status bar's negotiated cipher, host-key algorithm and key/credential name | ◆ **Shipped, on both surfaces, with three honest deviations.** `ISshConnection` and `ISftpSession` now both carry `Cipher` — the server-to-client algorithm off SSH.NET's own `ConnectionInfo.CurrentServerEncryption`, captured once at construction because a rekey is not an event SSH.NET raises — and `TerminalWorkspace.GetSessionFacts` hands the cipher and the host key's algorithm back to the shell the moment a session opens; `VaultViewModel.TryBuildAuthentication` now threads the authenticating key's or credential's own `Label` into `HostAuthentication.IdentityLabel`, all the way to `MainWindowViewModel`'s surface-aware `SessionCipher`, `SessionHostKeyAlgorithm` and `SessionIdentityLabel`, composed into one `SessionIdentityText` run for the status bar. Three deviations from the mock, not omissions: the algorithm prints exactly as negotiated (`ssh-ed25519`), not the design's shortened `ed25519`, because trimming it would be an edit to a string this client did not choose; the run is plain text rather than the design's clickable element, because there is no pin-details modal for a session that is already open, and drawing a click target for a screen that does not exist would itself be a fabrication; and a typed-password session — nothing filed in the keychain to name — shows the host-key algorithm alone, with no `·` after it, because there is no item behind the dot. |
+> | S3 dimmed in the design's own switcher | **Enabled.** The mock leaves S3 as future work; this application already has bucket browsing, so SSH, SFTP and S3 are a true three-way segment, wired to `IsSshShowing`, `IsTransfersShowing` and `IsBucketsShowing` exactly alike. |
+> | The S3/Buckets screen | **Did not get the session shell in v5b.** `TransfersScreen` serves both SFTP and S3 today and only the SFTP usage in `MainWindow.axaml` sat inside the new tab row/header/status bar/sidebar; the S3 usage was unchanged at the time. **v5c gives it the shell's own look without the machinery** — a 26-pixel padded, bordered, radius-12 container and nothing past that, since a bucket has no tab to close, no host to head a card with and no pin for a sidebar to show; see the v5c section, below. |
+> | No pins destination in the design at all | **The rail agrees with the design now.** `KnownHostsScreen` is still built and still reachable — from **Host keys** on the Keys screen's own header, which was always the second way in — but the rail's Pins row is gone. It was kept through v5b on the grounds that the mock has no screen for approved host keys, which is a reason for the screen to exist and was never a reason for a rail entry once the keychain had a door to the same place. Two rail rows landing on one screen is a rail that has to be read twice. |
+> | The popover's Settings and Preferences rows, and the design's own Settings-* family of screens | **Landed in v5c.** What was two doors to one room in v5b — Settings and Preferences both opening the same bare `Preferences` screen — is now two of three doors onto their own settings pages: Settings opens General, Preferences opens Preferences, and a third row, Vaults, opens Vaults. All three are real, distinct pages inside one settings mode; see the v5c section, below. |
+> | `· Org` after the user chip's name, and a `Primary` tag on a vault row in the popover | Neither. There is no organisation concept behind a vault — only the vault itself — and no vault is distinguished as primary; the popover's vault rows are the existing shown-vaults toggles, restyled. |
+> | The design's titlebar, which has nowhere for a sync indicator | `SYNCED` stays, on the titlebar's right side, ahead of the window's own minimise/maximise/close buttons — the one thing this titlebar keeps that the design's own does not draw at all. |
+> | Per-tab SFTP sessions, implied by a tab row shared between the terminal and the SFTP surface | **Not built, and not what shipped instead.** A click on the SFTP tab row runs `MainWindowViewModel.SelectFilesHostAsync`, which opens (or reuses) a second, SFTP-specific connection through the same "Browse files" plumbing a pin click already used — an honest second login, not a channel multiplexed onto the terminal's. `SelectedTab` moves with the click, which is also what keeps the sidebar's QUICK ACCESS in step — that list is keyed to `SelectedTab` on both surfaces, so selecting a different terminal tab afterwards can leave QUICK ACCESS naming a host that is not the one the remote pane is actually browsing. |
+> | A scroll-to-section affordance behind the sidebar's **+ Pin folder** row | Not built — this application has no way to open the host editor already scrolled to one card inside it. `PinFolderFromSidebarCommand` opens the same three-card editor the hosts screen's own EDIT reaches, at the top, which is the closest honest affordance rather than a new one. |
+> | The TRANSFERS strip's aggregate throughput readout (`8.4 MB/s` beside an arrow) | Omitted. `TransferRowViewModel.Progress` computes a rate per transfer; nothing sums those into one number for the whole queue, and inventing one would be exactly the fabricated fact this document's honesty rule forbids. |
+> | Freshness/"hot row" colouring and a selection ring on a finished transfer | Not drawn — there is no tracked notion of how recently a row finished, so nothing distinguishes a transfer that just completed from one that finished an hour ago. |
+> | The design's Material-Symbols-only "draft" glyph for a file row | `insert_drive_file`, the closest classic Material Icons has — a plain document rather than a page with a folded corner. Recorded as a substitution rather than silently swapped; "draft" is simply not a glyph this font contains. |
+> | `REMOTE` across the SFTP panes' arrow column | `HOST`/`BUCKET`, kept from before this pass rather than adopted from the design. The pair is the difference between a directory tree and a flat namespace with inferred folders, which is a distinction worth keeping even where the design collapses it. |
+> | `THIS MACHINE` on the local pane | `LOCAL`, the design's own word, adopted — nothing in this file's own comments ever defended "THIS MACHINE" the way HOST/BUCKET is defended above, so there was no reason to keep it. |
+> | The Keychain screen's `MODIFIED` column and `FINGERPRINT` box | Both dropped. `VaultItem` carries no timestamp of any kind, and `SshKeySecret` has no fingerprint field — computing one would mean parsing armour the type stores verbatim, an already-recorded gap this pass did not revisit. |
+> | Five separate add buttons (GENERATE / + SSH KEY / + PASSWORD / + TAG / + BUCKET) | One `+ New key` accent button, opening a menu of the same five actions in the same order — the design draws one button because its own mock has one "add" concept; this application has five, and none of them was dropped to fit. |
+> | `NEW ITEMS GO TO` on the vault-filing picker | `NEW ITEMS FILE TO`, the design's own wording, adopted — this screen's own comments already called the act "filing", so the label was this application's vocabulary already. |
+> | The Snippets screen's own on-screen name | **Snips**, matching the rail, which already said so. |
+> | A modified-date on a snippet card | Omitted — `SnippetSecret` carries no timestamp, the same absence every other item kind has. |
+> | The design's shorter captions — "Type into terminal", the rationale-only sentence for "runs on insert" | The screen's own longer, more actionable wording kept instead: the insert button still names the destination tab (`TYPE INTO {tab}` / `NO TERMINAL OPEN`), and the "runs on insert" caption still states the operational consequence rather than only the reason the setting exists. Both of the design's sentences are true; the ones already here say more. |
+> | A delete confirmation for a snippet | **Built, matching the mock.** `SnippetsViewModel.RequestDelete`/`ConfirmDelete`/`CancelDelete` say the same vault-wide-reach, tombstone, no-undo sentence `VaultViewModel.HowFarADeletionGoes` already says for every other item kind — additive beside the existing uncounted `DeleteCommand`, which the phone's own DELETE row still calls. |
+> | The Logs screen's own footer sentence about encryption | Adopted verbatim — verified against this screen's own header remark and ADR 0001 before shipping it, both true, so it is drawn as literal text rather than reworded. |
+>
+> ## The desktop's v5c — a settings area, at last
+>
+> An eighth pass, and the first to give the design's own Settings area a home of its own rather than folding
+> its two real screens into the rail's popover. Settings is a full-window **mode** now, swapped in wholesale
+> rather than laid over anything: its own 53px titlebar reading "Back to application", a 340px `SettingsNav`
+> rail — SETTINGS (General, Vaults, Account) above CUSTOMIZE (Security, Preferences, Groups, Tags), Logout
+> pinned below both — and a content column capped at the design's own 1100 pixels, which is 741 beside the
+> rail at the window's own minimum rather than the full 1100 (`LayoutHarness.SettingsContentWidth`). Seven
+> pages, not the design's eight: the ORGANISATION section and its one page are refused outright, below. Two
+> of the seven are wholly new — Account and Security did not exist as screens before this pass — and the
+> other five are the old `PreferencesScreen.axaml` and `VaultsScreen.axaml` split apart and restyled into the
+> design's own card idiom; both files are deleted, and every command either one offered is reachable here
+> exactly once. The popover's own Settings, Vaults and Preferences rows, and everywhere else in the codebase
+> that used to navigate to the bare `Preferences` or `Vaults` screen, now land in this mode instead —
+> `MainWindowViewModel.ShowScreen` redirects at that one point rather than at every caller that used to reach
+> either screen directly.
+>
+> The importer moved inside this same chrome rather than staying a screen of its own: it draws over the
+> Preferences page as a boolean overlay — `SettingsNav` stays lit on Preferences the whole time it is open,
+> and the titlebar's own back button reads "Back to preferences" instead of "Back to application" — restyled
+> to the design's own table (ALIAS/HOSTNAME/USER/PORT/WHAT THIS MEANS, one header tick-all button in place of
+> the old TICK ALL/TICK NONE pair). Known Hosts was restyled the same pass but kept its own place: it stays a
+> main-chrome screen, reached from the rail's Keys entry and now also from Security's own "Approved host
+> keys" row, because the design draws no page for it at all. And the S3/Buckets surface — which got none of
+> v5b's session shell — picked up that shell's *look* this time without its machinery: a 26-pixel padded,
+> bordered, radius-12 container and nothing else, since a bucket has no tab row, no host to head a card with,
+> no status-bar fact to print and no pin for a sidebar to show.
+>
+> | v5c element | What ships instead |
+> | --- | --- |
+> | The Settings-Organisation page, and the rail's own ORGANISATION section | Refused outright. No organisation entity exists anywhere in this product — a team is the membership list behind a shared vault, and there is exactly one tenant per deployment — so the rail simply has no third section; see `SettingsNav.axaml`'s own remark. |
+> | General: the update-channel switcher, Launch at login, Reopen tabs, the theme control and the language picker | Refused, and carried on the page's own NOT BUILT YET card rather than left silently missing: which channel a copy follows is fixed when it is built, nothing registers this application with Windows' own startup list or remembers a tab list across a launch, only the one dark theme exists, and there is no i18n anywhere in this client. |
+> | Vaults: the VAULT DEFAULTS card (auto-lock, require-password-on-unlock, relay) and the RECOVERY card (kit, export) | Both refused outright — none of the three settings exists, sync is always on, and there is no recovery kit and no export. |
+> | Vaults: "Manage devices", and the "3 devices" count on the sync line | Refused. There is no list-devices endpoint anywhere in this client. |
+> | Vaults: per-vault UNLOCKED/LOCKED chips, and the "Unlock ⟨vault⟩" modal | Refused. This application locks the keychain as a whole, not one vault at a time, so there is no per-vault state for a chip or a modal to act on; the keychain-level fact is the sync card's own dot instead. |
+> | Vaults: the magenta "Default" badge | Dropped rather than faked. The app does track a "new items go to" vault, but that lives on a different view model than the row being drawn here, and cross-referencing the two per card would be more machinery than the badge is worth. |
+> | Vaults: "Sorted by name ▾" | Decorative in the mock, and not drawn — `VaultsViewModel.Vaults` is already ordered personal-first-then-name, and there is no second order to switch to. |
+> | Vaults: member avatar stacks on every card | Refused on an unselected card — `VaultRowViewModel` carries a member *count*, not the members themselves, and only the selected vault's own member list is actually loaded. A count stands in on the card; avatars are real in the members panel, where the data is. |
+> | Account: Edit profile, Change passphrase, the DEVICES list, Sign out everywhere | All four refused — no endpoint exists for any of them, and the passphrase is a one-time enrollment choice with no change flow. What is new and real on this page instead: the SIGN-IN card's issuer sentence, `MainWindowViewModel.Issuer` off `MeResponse.Issuer`, cached the same turn the name and email are and surfaced here for the first time — "Signing in proves who you are — it never decrypts a vault." |
+> | Security: the strict-host-key toggle, the allowed-algorithms chips, the auto-lock link row | Refused — the app always asks on a changed key, there is no algorithm allow-list anywhere in the SSH stack, and there is no auto-lock setting for the link to point at. |
+> | Security: RECENT SECURITY EVENTS as a list of rows | Shipped as the link alone, "See all activity in Logs." A real list would mean a filtered read over `LogsViewModel`'s two logs, and which entries count as "security" is a judgement call the design does not resolve; the honest link is complete on its own. |
+> | Preferences: the device-name editor, Font/Cursor/Scrollback/Copy on select/Terminal bell, the clipboard-clear delay, confirm-run-on-insert | All refused, on the page's own carried-over NOT-BUILT idiom — none of the seven has anything behind it. Windows Hello's register/"Stop unlocking here" pair moved off this page to Security instead, one home rather than two. |
+> | Groups: drag-to-reorder, and the mock's "the order here is the order there" sentence | Refused. `HostGroupRowViewModel` orders by the vault new items go into, then by vault name, then by label — there is no manual order to drag into. The page prints the true sentence instead and draws no drag handles. |
+> | Tags: the LAST APPLIED column | Refused — no timestamp of when a tag was last put on a host exists anywhere in this client. |
+> | Tags: a per-tag vault chip | Refused — `TagRowViewModel` carries a label and a host count and nothing else; there is no vault id on the row for a chip to read. |
+> | Import: the "saving to ⟨vault⟩ ▾" picker | Refused as a control. Every import writes through the same `session.ActiveVaultId` every other bulk write does, so there is no per-import target to choose — the footer prints the vault's name as a fact, `ImportViewModel.SelectionSummary`, rather than as a `▾`. |
+> | Import: the old ADDRESS/AUTHENTICATION/STATE columns | Folded into the design's own column set — ALIAS/HOSTNAME/USER/PORT — with the authentication text moved onto the alias cell's own tooltip rather than kept as a column of its own. |
+> | Import: TICK ALL / TICK NONE | Replaced by the design's own header tick-all button, one press that ticks or unticks every row through `ImportRowViewModel.ToggleAllCommand`. |
+> | Import as a nav-rail destination | Refused. It is a boolean overlay over the Preferences settings page (`MainWindowViewModel.IsImportOpen`), not a page of its own — `SettingsNav` stays lit on Preferences throughout, exactly as `Import.dc.html` draws it. |
+> | The design's fixed 1100px content column | `MaxWidth="1100"` rather than a fixed width — the column is 741 pixels beside the 340px rail at the window's own minimum, where a fixed 1100 would not fit. |
+> | The status bar and the update banner, while in settings mode | Both hidden — `MainWindow.axaml`'s bottom two rows read `!IsSettingsMode`. The design's own settings titlebar has no room for either, and there is nowhere honest to draw them instead. |
+> | The design's sentence-case button copy | Not adopted. Every button on every settings page keeps this application's own ALL-CAPS mono convention — `CHECK NOW`, `SIGN OUT`, `OPEN IMPORTER` — over the mock's own "Check now". |
+> | The design's assumption that Settings is the only thing on screen | Not followed. The quick-connect palette and an unapproved host key's decision card both still draw over settings mode exactly as they interrupt every other screen — a connection question does not stop mattering because the window happens to be showing Settings; see `MainWindow.axaml`'s own remark on the ordering of its Panel. |
+> | The importer's own sentence, "This is the only control in DodoSSH that opens key material from a directory you did not point at file by file — nothing is read until Import is pressed." | Adopted verbatim — verified against `ImportViewModel.ScanAsync` and `ImportAsync`: scanning never calls `SshConfigLocator.ReadIdentity`, and `ImportAsync` is the only path that ever does. |
+> | The tags page's own sentence, "Renaming here is one write and every host follows." | Adopted verbatim — verified against `SaveTagAsync`: a host names a tag by id, never by label, so a rename touches nothing but the tag item itself. |
+> | The known-hosts screen's own intro, "Every pin is a decision recorded at the moment of connecting. Fingerprints are never trimmed — compare them character by character against what the operator published." | Adopted verbatim, matching the stance the rest of this application already takes on a pin. |
 
 Most of it landed. This file is the rest: every element of that design with nothing behind it, which
 project each piece would have to land in, and **what the shipped interface does instead**. That last
@@ -173,8 +384,9 @@ it for revocation, for relay session recording and for what locking does not do
 rendered with invented data to make a screenshot look finished. Where a feature does not exist, the
 interface either omits it or says so.
 
-Read alongside [the milestone plan](../README.md#milestones). Most of what follows is not a defect; it is
-M2 and M3 arriving in a design before it arrives in the code.
+Read alongside [the milestone plan](../README.md#milestones). Most of what follows is not a defect. M2 and
+M3 have since arrived and their rows say so; what is left is either M5 — per-item content keys, and the
+sharing that would rest on them — or a decision recorded here not to build the thing at all.
 
 ## The shape of the problem
 
@@ -204,9 +416,12 @@ reason and is gone: membership is granted only to an account that exists, so the
 reservation fails to cover. The status stays reserved, because the column holds it in nobody's database and
 a client must not fail on a value a later server might send. See [ADR 0009](adr/0009-team-access-model.md).
 
-**The client has no preferences store.** It writes exactly two files — `cache.db` and `device.key` — and the
-cache has six tables, none of them settings. Nothing on the design's TERMINAL preferences panel can be
-saved, and there is no frame on the terminal data plane that would carry a change to the renderer anyway.
+**The client's settings file holds two things.** It writes three files — `cache.db`, `device.key` and
+`settings.json` — and the cache's six tables still hold none of them. `ClientSettings` carries the
+terminal's text size and whether this machine looks for a newer build on its own, and both survive a
+restart. Everything else on the design's TERMINAL panel is unbuilt rather than unbuildable: the data plane
+grew `TerminalServerOpcode.FontSize` to carry the one that shipped, so a second option is an opcode, a
+setting and a control rather than a new mechanism.
 
 **Three things the design did not ask for and this build now has.** A key can be generated in the client
 rather than pasted in (`SshKeyGenerator`, and the `openssh-key-v1` container is written by hand — see
@@ -297,7 +512,7 @@ field cannot be removed and stays as a permanently refused member; `SyncEndpoint
 | Status bar port forwards | client-ssh | Port forwarding. See below. | Omitted. |
 | Status bar `sftp · 2 transfers` | client-app | Nothing now — file transfer is built. What is missing is the count reaching the status bar, which is a screen away from where the queue lives. | Omitted from the status bar. The queue itself is on the FILES screen, with a row per transfer. |
 | Status bar `locks in 09:41` | client-app | An idle auto-lock. See preferences below. | Omitted. |
-| IBM Plex Mono / IBM Plex Sans | ui | Shipping the font files as `AvaloniaResource` and registering them. The design loads them from Google Fonts, which a desktop app cannot. | Inter (already embedded) for prose, and the system monospace stack the terminal already names. Named once in `App.axaml` as `MonoFont`, so the substitution is reversible in one place. |
+| IBM Plex Mono / IBM Plex Sans | ui | Shipping the font files as `AvaloniaResource` and registering them. The design loads them from Google Fonts, which a desktop app cannot. | **Shipped, since v5, and not IBM Plex.** The v5 design asked for Montserrat, JetBrains Mono and Material Icons instead, and all three are embedded as `AvaloniaResource` in the Shell project rather than requested by name and left to whatever the machine happens to have. Montserrat is the default sans, with Inter — the earlier substitution — kept only as its fallback; `MonoFont` puts JetBrains Mono first, ahead of the system-monospace stack the terminal still falls back to. Named once each in `Program.cs` and `Palette.axaml`, so a later design's fonts are still a one-line swap. |
 | `⌘K`, `⌥↵` | ui | Nothing; the design is Mac-flavoured. | `CTRL K`. Development is Windows-only today (`docs/platform-flags.md`). |
 
 **Also worth knowing:** custom chrome went in as `SystemDecorations="BorderOnly"`, not by extending the
@@ -322,11 +537,11 @@ caption buttons and window title drawn on top of the application's own — two s
 | `SNIPPETS` panel, `↵` to run | client-domain | A snippet item type (`Snippet = 8`, reserved). | **Shipped**, as a screen rather than a panel. `↵` is per snippet and off by default: inserting types the command at the prompt and stops, because nothing here can tell whether the terminal is at a prompt at all. |
 | Broadcast to all panes (`⌥↵`) | client-ssh | Input is routed strictly by session id in `TerminalDataPlane.Dispatch`; there is no fan-out. Needs splits first. | Omitted. |
 | Pane header `24ms` | client-ssh | Round-trip measurement. SSH.NET offers no RTT API. | Omitted. |
-| Pane header `aes256-gcm` | client-ssh | **The closest miss on this list.** `SshNetConnection` holds the `SshClient`, so `ConnectionInfo.CurrentServerEncryption` is right there — it just is not on `ISshConnection` or surfaced by `TerminalWorkspace`. | Omitted; the tab strip shows the account and endpoint actually dialled. |
+| Pane header `aes256-gcm` | client-ssh | **No longer a miss — the plumbing this row used to lack now exists.** `ISshConnection.Cipher` and `TerminalWorkspace.GetSessionFacts` were built for the v5b session shell's status bar; see that row in the v5b section, above. | Still not on a pane header — this pre-v5b element does not exist as its own piece of chrome any more. The fact it wanted to show is drawn instead where v5b moved it: the status bar's `SessionCipher`, beside CONNECTED. The session shell's host header separately shows the account and endpoint actually dialled (v3–v4: the tab strip did; v5b moved that fact to `SessionHeader`, off `MainWindowViewModel.SessionAddress`, when the strip was retired). |
 | Pane header showing the running command and `following` | client-ssh | The host moves opaque bytes and never parses terminal output. Would need shell integration (OSC 133) on the remote. | Omitted. |
 | A `local · zsh` tab | client-ssh | Every session here is an SSH channel. Needs ConPTY and a second session kind. | Omitted. |
 | Tab strip `+` button | ui | Not missing so much as redundant: the real operation is *select a host, press Connect*, which the hosts grid already is. | **Shipped**, as the palette rather than a menu: it opens what Ctrl+K opens, so the strip and the shortcut are one way of doing one thing. A `MenuFlyout` offering "SSH" and "local shell" is the nicer answer and is not verifiably safe above the terminal's native child window — and there is no local shell to offer. |
-| Terminal font size (`--termfs`, 11–16px) | client-storage | See preferences. | Fixed at the renderer's 13px. |
+| Terminal font size (`--termfs`, 11–16px) | client-storage | — | **Shipped**, and wider than the design's range: 8–32px from the preferences screen or Ctrl+plus/minus/0 over a terminal, carried by `TerminalServerOpcode.FontSize` and kept in `settings.json`. It resizes the grid rather than magnifying it, so every live pane refits and the remotes are told how many columns they now have. |
 
 ---
 
@@ -350,11 +565,11 @@ connect, and it goes through the same host key gate, the same pin and the same t
 | Remote listing with `NAME/SIZE/MODIFIED/PERMS` | client-ssh | All four. `PosixMode` renders `drwxr-xr-x` from the bits SFTP hands over; setuid, setgid and sticky are not shown, because SSH.NET does not surface them and `rwx` where `rws` is true would be worse than nothing. |
 | Local listing | client-transfer | `LocalDirectory`, which is where this client's `System.IO` now lives. `PERMS` is blank on the local side rather than filled with a plausible-looking POSIX mode that is not a fact about a file on Windows. |
 | Transfer queue, progress, throughput | client-transfer | `FileTransferQueue`. One transfer at a time, so the rate on a row is the rate of the link rather than a share of it. Throughput is measured over a half-second window, not averaged since the start. |
-| `resume supported` | client-transfer | **Within a run of the application.** Every transfer writes to a `.dodossh-part` file beside its destination and is renamed into place at the end, so an interrupted one can never be mistaken for a finished one, and `RESUME` carries on from the part file's own length. A part file found at startup is *not* resumed: nothing records what wrote it, and resuming on the strength of a name matching is how a corrupt artefact gets delivered with nothing reporting a failure. Making it survive a restart needs the preferences store this client has not got — see below. |
+| `resume supported` | client-transfer | **Within a run of the application.** Every transfer writes to a `.dodossh-part` file beside its destination and is renamed into place at the end, so an interrupted one can never be mistaken for a finished one, and `RESUME` carries on from the part file's own length. A part file found at startup is *not* resumed: nothing records what wrote it, and resuming on the strength of a name matching is how a corrupt artefact gets delivered with nothing reporting a failure. Making it survive a restart needs somewhere to record *what* wrote each part file — the source, the offset and the run — which is a table rather than a setting; `settings.json` holds two scalars and is not that. |
 
 | Design element | Layer | What it would take | What ships instead |
 | --- | --- | --- | --- |
-| Per-host last directory | client-storage | Somewhere to persist two panes' navigation state. There is still no settings table. | The remote pane opens on the account's home directory, which the server canonicalises during the handshake; the local pane opens on the user profile. |
+| Per-host last directory | client-storage | Somewhere to persist two panes' navigation state per host, which is a table rather than a scalar — `settings.json` holds two of those and is the wrong shape for a row per machine. | The remote pane opens on the account's home directory, which the server canonicalises during the handshake; the local pane opens on the user profile. |
 | `sftp over bastion-eu` | client-ssh | Jump hosts, as above. `HostSecret.JumpHostIds` is still stored, synced, merged and read by nothing. | Omitted. |
 | Overwriting a file that is already there | ui | A prompt, which means a modal this window has no idiom for. | Refused, with the name that is in the way. The remote pane has DELETE and MKDIR so the refusal is not a dead end. |
 | Dragging between the panes | ui | Drag-and-drop between two `ListBox`es, plus a drop target that is a directory rather than a row. | Two arrow buttons between the panes, pointing at the pane the file is going to. |
@@ -476,8 +691,8 @@ lists the rest as absent rather than omitting it silently.
 
 | Design element | Layer | What it would take |
 | --- | --- | --- |
-| Any preference, saved | client-storage | A seventh table in the client cache, or a preference item type in the vault. Which one matters: the vault syncs, so a preference stored there follows you between machines, and a terminal font size probably should not. |
-| Terminal font, size, cursor style, cursor blink, scrollback | client-terminal | The store, **and** a frame to carry it. `TerminalServerOpcode` has four values (`Output`, `SessionOpened`, `SessionClosed`, `SessionActivated`) and none carries an option. |
+| Any preference, saved | — | **Two are.** `settings.json` beside the cache holds the terminal's text size and whether this machine checks for updates on its own. It is deliberately not in the vault: the vault syncs, so a preference kept there would follow you between machines, and a font size chosen for a laptop screen should not arrive on a desktop. Anything per-host or per-item wants a table instead — see the transfers rows above. |
+| Terminal font, cursor style, cursor blink, scrollback | client-terminal | A setting, an opcode and a control, in that order — the mechanism is no longer the obstacle. `TerminalServerOpcode.FontSize` proved the path and `ClientSettings` is where the value would live; what none of these four has is somebody deciding it is worth a row on the preferences screen. |
 | Font family picker offering IBM Plex Mono | ui | Shipping the font, as above. |
 | Audible bell | client-terminal + ui | xterm.js 6 removed `bellStyle` and `bellSound`; what remains is an `onBell` event, so the page would have to make the sound itself. |
 | `GENERAL` section | ui | There is no general setting to put in it. The theme is fixed by decision, and window size is not persisted. |
@@ -505,3 +720,48 @@ grid of cards with a drawer — see above. The split it describes did not change
 running, so a tab list rebuilt per unlock would lose track of sessions that are still connected — the very
 sessions the unlock screen already counts. `TerminalWorkspace` gained `SessionActivated` on the wire,
 `IsSessionLive`, and a `SessionEnded` event so a tab can stop claiming to be connected.
+
+---
+
+## v5c-4 — two more, asked for after living with v5b
+
+**The session shell's host header is gone, and it is a deliberate departure from the design.**
+`Terminal.dc.html` and `SFTP.dc.html` both draw a 60-pixel row above the pane carrying the address on the
+left and a cross-surface button on the right, and v5b shipped it as `SessionHeader.axaml`. Both of the two
+facts it held now live at the head of the sidebar beside the pane — the address as its own line, and the
+button stretched across the column under it — and the pane is 60 pixels taller for it. The reasoning is the
+one the design cannot see from a mock: this is a window somebody keeps a terminal open in all day, and a
+full-width strip repeating an address the tab already names was the cheapest 60 pixels in the layout to give
+back. `LayoutHarness.SessionScreenHeight` no longer subtracts a header, for the same reason it stopped
+subtracting the retired window-wide tab strip.
+
+**The sidebar closes, which the design has no state for.** 300 pixels of a 1081-pixel minimum is a lot to
+spend on a list that is often two rows long, so `MainWindowViewModel.IsSessionSidebarOpen` folds the column
+to a 34-pixel rail carrying the chevron that brings it back — a rail rather than nothing, because a panel
+that vanishes without trace is one people report as lost. The choice is written through to
+`ClientSettings.SessionSidebarOpen` rather than held for the session: it is a decision about how much of the
+window a terminal gets, and one that had to be made again on every launch would not really be on offer.
+
+## v5c-5 — one door to settings, not three
+
+**The user-chip popover keeps its Settings row and has lost Vaults and Preferences.** v5c gave it three, and
+the table above records why: until that pass "Settings" and "Preferences" had been two doors onto the same
+bare screen, so splitting them was a real fix, and Vaults arrived beside them when `VaultsScreen` became a
+settings page too.
+
+Living with it made the shape of the mistake clear. All three opened the same full-window mode and differed
+only in which of its seven pages they landed on — and `SettingsNav`, which lists every one of those pages
+including Vaults and Preferences, is on screen the entire time you are in there. So two of the three rows
+were deep links into a place the third one opens, saving one click and costing a menu that read as though it
+offered three destinations. A menu at the foot of the rail is for getting into settings at all; which page
+you want is a question that has a rail of its own once you are there.
+
+Nothing became unreachable: both pages are one click deeper, from the rail that was always going to be the
+way to them. `NavRailTests` asserts the rows are gone from the markup rather than merely unwired, because a
+row left behind with no handler is a menu entry that does nothing when pressed — which reads as a broken
+application rather than a tidied menu.
+
+Logout stays, and still lands on a page rather than on General: its confirmation card is drawn inline on the
+Account page and nowhere else, so arming it from anywhere else would be a card raised where nobody could see
+it. The vault switches and **New vault** above the divider stay too — they are what the popover is for
+besides settings, and neither is a page of the settings mode.

docs/platform-flags.md

@@ -3,8 +3,18 @@
 Things known or suspected to behave differently outside Windows, plus deployment gotchas that
 have already cost time once. Development is Windows-first, but **the full test suite now runs on
 Linux in CI on every change**, so a Linux claim here is usually a measurement now rather than a
-suspicion. **macOS is still untested**, and anything marked *unverified* has not run on the platform
-in question and must not be assumed to work.
+suspicion. Anything marked *unverified* has not run on the platform in question and must not be
+assumed to work.
+
+**macOS now builds and packages, and has still never run.** The distinction matters more here than
+anywhere else on this page, because the two halves are verified in completely different places. The
+build is measured on every main and tag build: CI publishes `osx-arm64` and runs `vpk [osx] bundle`
+on a Linux runner, which is enough to catch a restore graph with no macOS native asset and an `.app`
+that will not compose. Everything past that — whether the window draws, whether the terminal's
+loopback WebSocket reaches WKWebView, whether the Secure Enclave holds a device key — is verified
+only by a person walking Phase 18 of [manual-checks.md](manual-checks.md) on a Mac, because **there
+is no macOS runner in CI**. Treat every macOS runtime claim below as unverified unless it says
+otherwise.
 
 Each entry says what the risk is, why it matters, and what to do about it. Delete an entry when it
 has been verified or made moot — not when it merely stops being convenient.
@@ -17,6 +27,19 @@ docs/crypto.md §1. *Already mitigated* — but if a BCL AEAD path is ever added
 **must** gate on `IsSupported` rather than assuming availability, or the client will fail to open
 any vault on macOS.
 
+**The Secure Enclave holds P-256 keys and nothing else**, which is why `MacDeviceKeyStore` wraps the
+device key with ECIES rather than with the RSA-OAEP the Windows store uses. It will not hold an RSA
+key at any size, so this is not a preference. The useful consequence is that the macOS shape is
+*better* than the Windows one: `SecKeyCopyPublicKey` works on an enclave key without prompting, so
+registering a device is silent and only unlock asks — where Windows raises a dialog at key creation
+too. *Unverified:* no enclave call in this repository has ever run.
+
+**Three ordinary Macs have no usable enclave**, and `IsSupported` probes rather than infers for that
+reason: an Intel machine without a T2, a machine with no login password set, and — the one that
+surprises people — **any build that is not code signed**, because enclave key creation needs a
+signing identity. So `dotnet run` correctly offers no device key at all. Do not "fix" this by
+checking the OS instead; the offer would then put a wrap on the server that nothing can ever open.
+
 **Argon2id timings are measured on one Windows machine only.** 256 MiB with t=4 took 323 ms here.
 The floor and ceiling in `EnrollmentLimits` were chosen against that number. *Unverified
 elsewhere:* recalibrate on the slowest target platform before recommending a default profile,
@@ -26,7 +49,11 @@ and the parameters are stored per user at enrollment, so a bad default is a per-
 **libsodium ships native binaries per RID.** This complicates single-file and AOT publishing, and
 on macOS every native library (`libsodium`, `libSkiaSharp`, `libHarfBuzzSharp`, `libe_sqlite3`)
 must be signed **individually** with `--options runtime --timestamp` before the bundle is signed,
-or notarization fails with an error that does not name the offending file.
+or notarization fails with an error that does not name the offending file. *Mitigated* in
+`scripts/release-macos.sh`, which signs every `.dylib` and `createdump` in a loop before vpk touches
+anything — vpk's own pass uses `codesign --deep`, which is the shape Apple documents as wrong for
+nested code and is the likeliest source of that unnamed rejection. The loop looks redundant next to
+`--deep` and is not; do not delete it because a release once succeeded without it.
 
 ## Desktop client
 
@@ -361,11 +388,55 @@ agent of our own plus ProxyJump covers the real use cases.
 **The SSH suite pulls `linuxserver/openssh-server` from Docker Hub**, which is rate-limited for
 unauthenticated pulls. If CI starts failing on image pulls rather than on tests, that is why.
 
+**That suite has an intermittent `The connection was closed by the remote host`**, on whichever test
+connects first, within tens of milliseconds. Seen in CI and reproducible locally. *Mitigated, not
+solved:* `SshServerFixture` now raises sshd's `MaxStartups` from its compiled-in `10:30:100`, which
+refuses connections at random past ten unauthenticated ones in flight — reachable because xUnit runs
+test classes in parallel and most of them connect. The fixture comment carries the full argument and
+is explicit that the cure is unproven.
+
+**And the reason it is unproven is a measurement trap worth not falling into twice.** Docker
+throughput on the Windows development machine swings enough to swamp the effect: the identical
+unmodified suite ran 85/85 clean and, an hour later, failed 13 runs out of 15. Any before/after flake
+comparison taken there is noise. Measure this class of thing in CI, or make the server say why —
+raise sshd's `LogLevel`, disable Ryuk so the container outlives the run, and read `docker logs`.
+
 **MSIX packaging is ruled out, not merely deprioritised.** A packaged app runs WebView2 in an
 AppContainer where loopback connections are blocked without a `CheckNetIsolation` exemption. The
 terminal data plane *is* a loopback WebSocket, so MSIX would break the product outright. Velopack
 for Windows/macOS/AppImage; Flatpak and deb/rpm defer updates to the package manager.
 
+**The App Sandbox is ruled out on macOS for the same reason, and the entitlements say so.** A
+sandboxed process cannot listen on loopback without `com.apple.security.network.server`, and the
+terminal is that listener. Developer ID distribution outside the App Store does not require the
+sandbox, so this costs nothing today — but it does mean the Mac App Store is closed to this
+application without solving the data plane differently first. See
+`build/macos/DodoSSH.entitlements`.
+
+**The hardened runtime is not optional and .NET needs four holes punched in it.** Notarization
+refuses a Developer ID submission without it, and CoreCLR will not start under it without
+`allow-jit` and `allow-unsigned-executable-memory` — both, not either, because the runtime allocates
+executable memory outside the `MAP_JIT` path as well. `disable-library-validation` and
+`allow-dyld-environment-variables` are needed for Velopack's updater rather than for the runtime.
+Each is argued individually in the entitlements file; the failure mode for a missing one is a
+process that dies during runtime initialisation, before anything exists that could report it.
+
+**`vpk` cross-compiles to macOS only as far as the bundle.** `vpk [osx] bundle` runs anywhere and
+produces a real `.app`; there is no `[osx] pack` off a Mac, because pack drives `codesign`,
+`notarytool` and `stapler`. So CI can prove the bundle builds and only a Mac can produce something
+installable. Note this is the *opposite* of the Windows story, where `vpk [win] pack` builds the
+whole installer on Linux — the asymmetry is Apple tooling, not a Velopack limitation.
+
+**A custom `Info.plist` is copied verbatim by vpk, with no substitution whatsoever.** That is why
+`--plist` and `--bundleId` are mutually exclusive, and why `build/macos/Info.plist.template` is a
+template the release script renders rather than a committed file. A committed plist would carry one
+version into every release afterwards, and the symptom is silent: Velopack's index would still be
+right, the updater would still work, and only Get Info and any crash report would disagree.
+
+**macOS app icons live on an 824-in-1024 grid.** An icon that bleeds to the edge of its canvas is
+not bolder, it is the one icon in the Dock that is too big. `dodossh-icon.ps1` draws the `.icns` at
+that fraction and the `.ico` at full bleed, from one geometry.
+
 *Checked rather than assumed, now that Velopack is actually wired up:* its Windows path does not
 reintroduce the thing MSIX was ruled out for. `Setup.exe` is an ordinary Win32 executable that unpacks a
 directory under `%LOCALAPPDATA%` and creates shortcuts — there is no `AppxManifest`, no package identity,
@@ -595,9 +666,23 @@ targeted, not which RID. Only signing needs Windows tooling, which is why `ci.ym
 without a target *evaluates* the project and runs nothing, while MinVer computes the version inside a
 target — so the read comes back as the SDK default on a full checkout with every tag present, which looks
 exactly like a version that was never configured. `-t:MinVer` makes `-getProperty` report the value after
-that target has run, and both readers of it — the tag check in `ci.yml` and `scripts/release-windows.ps1`
-— pass it. Neither had, and neither had ever run: the CI check is `if:` a tag ref and there are no tags
-yet, so the first release would have been refused by its own guard, which would have blamed `fetch-depth`.
+that target has run, and every reader of it — the tag check in `ci.yml`, the desktop nightly job, and
+`scripts/release-windows.ps1` — passes it. Neither of the first two had, and neither had ever run: the CI
+check is `if:` a tag ref and there are no tags yet, so the first release would have been refused by its own
+guard, which would have blamed `fetch-depth`.
+
+**And naming that target requires a restore first, which is a second failure wearing a very different
+face.** MinVer arrives as a package, so its target is imported from `obj/*.nuget.g.targets` and does not
+exist at all on a clean checkout:
+
+```
+error MSB4057: The target "MinVer" does not exist in the project.
+```
+
+That reads like a typo in the workflow rather than like a missing restore, and it does not reproduce on any
+machine that has built the project before — which is every developer machine and no fresh runner. The
+build job's tag check is safe because it runs after that job's own restore; the desktop nightly job and the
+release script each restore before reading, deliberately and with a comment saying why.
 
 **The Android head's lock file is outside the solution, so nothing checks it until the android job runs
 — and the android job was broken for an unrelated reason for the whole of the release that went stale.**
@@ -627,6 +712,30 @@ The lasting hazard is the first paragraph and not the fix. Any change to a share
 to a lock file this repository cannot verify from a machine without the Android workload, and it will go
 on being noticed later than every other one.
 
+**A lock file can go stale with nothing in this repository changing, because `Microsoft.NET.ILLink.Tasks`
+is versioned by the SDK and `global.json` lets the SDK float.** The reference is implicit — nothing in any
+`.csproj` asks for it — and its version tracks the runtime patch band, while `global.json` pins only
+`10.0.100` with `rollForward: latestMinor`. So `setup-dotnet` installs whatever the newest 10.x SDK is on
+the day, and the moment that SDK's band moves, locked-mode restore stops:
+
+```
+error NU1004: The package reference Microsoft.NET.ILLink.Tasks version has changed
+from [10.0.10, ) to [10.0.11, ).
+```
+
+It named `DodoSSH.Client.Android`, `DodoSSH.Contracts` and `DodoSSH.Crypto` — the three lock files that
+carry the entry — on a commit that touched none of them and no dependency at all.
+
+The fix is `--force-evaluate` on those three, **from a machine whose SDK is at least as new as the
+runner's**, which is the part that is easy to get wrong: a `--force-evaluate` from an older SDK rewrites
+the lock at the older version, changes nothing, and looks like it worked. Check `dotnet --version` against
+the version in the error before believing a regeneration.
+
+This will recur on every SDK patch that moves the band. It is the accepted cost of letting the SDK float:
+the alternative is pinning an exact SDK in `global.json`, which trades a recurring lock-file bump for a
+recurring toolchain bump and makes every contributor install one specific SDK. Neither is free, and this
+repository has chosen the floating side deliberately.
+
 **.NET for Android cannot be built on a musl host, and this project's runner is Alpine. Every message the
 toolchain produces on the way to saying so names a missing file that is present.** Three CI rounds went
 into this and the first two fixed symptoms, so the messages are worth reading in the order they arrive.

docs/reaching-a-host-you-cannot-dial.md

@@ -0,0 +1,314 @@
+# Reaching a host you cannot dial
+
+Some machines do not answer from where the user is sitting. This product has **two** answers to that, and
+one of them works.
+
+- The **relay** pipes raw TCP through the deployment. Both halves are built: the server issues a ticket and
+  pipes a WebSocket to the host, and the client reaches it through a loopback SOCKS5 bridge. Step 3 below.
+- A **jump host** reaches the target through a machine already in the keychain. `HostSecret.JumpHostIds`
+  stores the chain; nothing writes it and nothing reads it.
+
+◆ **An earlier version of this page said the relay's server half was "built and shipped". It was not.** What
+existed was ADR 0004, the contract records, the options and the CHECK constraint — and `RelayOptions` said,
+in its own remarks, that there was no endpoint and no dialler. Both halves were built together at step 3.
+
+This document is the comparison between them, which is the thing that has to be settled before either is
+built, and then the plan. It replaces an earlier draft of `docs/jump-hosts.md` that recommended deleting the
+jump chain — see the last section for why that was wrong.
+
+> **Status: steps 0, 1 and 3 done, and step 4 for the relay. Jump hosts are next.**
+>
+> | Step | State | Notes |
+> | --- | --- | --- |
+> | 0. Stop promising the relay | **Done, then retired at step 3** | The "not wired up yet" sentence came out of both heads when the relay started working |
+> | 1. Dial through a loopback proxy | **Done** | `SshLoopbackProxy` on the request, honoured in `BuildConnectionInfo`. **Not the bridge** — see below |
+> | 2. Jump hosts over it | Not started | No server change, and no bridge either: SSH.NET's own forward is the proxy |
+> | 3. The relay: the bridge, then the ticket | **Done** | Server: `Features/Relay`. Client: `RelayBridge`, `ServerRelay`, and `VaultViewModel.OpenThroughAsync`. See "What step 3 built" |
+> | 4. File transfer parity | **Done for the relay** | `TransfersViewModel` opens a relay for a relayed host exactly as the shell does. A jump chain will need the same line when step 2 lands |
+
+## ◆ The relay's checkbox is a false promise, and that is a defect
+
+`HostSecret.RelayEnabled` is stored, validated — a relay host may not inherit its port — encoded, merged,
+and drawn as a checkbox in the host editor on **both** heads. The desktop's says *"Connect through the
+server relay"* and warns underneath that the address will be stored on the server in plain text. The
+phone's says the same at more length.
+
+Nothing on the client reads it. `VaultViewModel` builds `SshConnectionRequest(hostname, port, username,
+credential)` and `SshNetConnectionFactory` dials that address directly, whether the box is ticked or not.
+
+So a user who ticks it **pays the privacy and gets nothing**: the host's address and port leave the
+encrypted payload and land in plaintext columns on the server — the one deliberate concession in the whole
+design, per ADR 0004 — and the connection is still made from their laptop to the machine they already could
+not reach. It then fails exactly as it did before, with no hint that the box did nothing.
+
+This is worse than the jump chain, which is invisible and harmless. It is a control that spends something
+real.
+
+**Step 0 is done.** Both heads now name the absence in the label and lead the paragraph under it with what
+ticking the box actually does today — the way this codebase already handles port forwarding on the phone's
+More screen. It is left *tickable* rather than disabled, deliberately: a host that already carries the flag
+has to be able to lose it, and a disabled control would trap the concession on. The sentence comes out again
+at step 3.
+
+## The comparison
+
+Both answers put something between the user and a machine they cannot dial. What differs is *what* is in
+between, what it costs, and who has to own it.
+
+| | Relay | Jump host |
+| --- | --- | --- |
+| **Reaches** | Anything the **deployment** can reach | Anything a **machine already in the keychain** can reach |
+| **Asks of the deployment** | It must sit where it can dial the target, and have the relay enabled | Nothing. The server is not involved at all |
+| **Tells the operator** | The host's address and port, in plaintext columns, for every opted-in host — plus an audit row per session: target, duration, bytes, close reason, client IP | Nothing beyond the sync metadata every item already produces |
+| **The intermediate's credentials** | None to manage. The deployment is the intermediate | The bastion is an ordinary host: its own key or password, its own host key to pin, its own group defaults |
+| **Where SSH terminates** | On the laptop. The relay sees ciphertext, and ADR 0004 is emphatic that no recording is possible | On the laptop. The bastion forwards a TCP stream inside a session the user opened to it |
+| **When it is unavailable** | Deployment down, no connection — including to hosts that were reachable directly | Bastion down, no connection to what is behind it |
+| **Fits an estate where** | The DodoSSH server is *inside* the network the targets are on | A bastion is the policy and the server is outside — which is the ordinary enterprise shape |
+| **Auditable by the operator** | Yes, coarsely, and that is a feature for a team deployment | No, and that is a feature for a private one |
+
+**The two are not substitutes, and the deciding question is where the deployment sits.** The relay only
+answers "unreachable" when the server has line of sight the laptop lacks — a deployment inside the VPC, on
+the office network, on the same Tailnet. Point it at a self-hosted box outside the target's network, which
+is what most people running this on a VPS will have, and the relay reaches nothing the laptop could not
+already reach.
+
+**And you do not get to choose other people's topology.** Shipping this to strangers means shipping into
+estates whose shape is already decided, and bastion-fronted is the common one. Their `ssh_config` says so:
+the importer reads `ProxyJump`, records it as an option and writes a note on the host saying *"DodoSSH does
+not route through a jump host yet"* — a first-run experience that names the limitation on the hosts it
+matters for.
+
+**A relay is not a bastion with better manners.** ADR 0004 rejected "server terminates SSH" and kept
+zero-knowledge, which is right and is not what a jump host asks for either: forwarding a TCP stream through
+a machine the user has authenticated to reveals nothing to the operator, because the operator is not in it.
+The privacy ordering is the opposite of what the ADR's framing suggests — the relay is the mechanism that
+costs a plaintext address, and the jump host is the one that costs nothing.
+
+## They are one piece of work, and ADR 0004 says so
+
+The last consequence in ADR 0004, written before either half was built:
+
+> On the client, SSH.NET cannot be handed a pre-connected stream, so the relay is reached via a loopback TCP
+> bridge. The same bridge provides ProxyJump via a SOCKS5 dynamic forward — **one mechanism, two features**.
+
+That is the plan, and it holds up against the pinned package. SSH.NET 2025.1.0 offers
+`ForwardedPortDynamic`, which is a SOCKS5 proxy served over an established `SshClient`, and
+`ConnectionInfo(host, port, username, ProxyTypes, proxyHost, proxyPort, proxyUsername, proxyPassword,
+AuthenticationMethod[])` with `ProxyTypes.Socks5` — checked in `Renci.SshNet.xml` rather than remembered. So:
+
+- **Jump host:** connect to the bastion as an ordinary host, `AddForwardedPort(new ForwardedPortDynamic(0))`
+  on it, then dial the target with a `ConnectionInfo` pointed at that loopback SOCKS5 port. A chain of two
+  is the same trick twice.
+- **Relay:** the same shape with a different thing on the loopback socket — a listener that pipes bytes into
+  the `dodossh.relay.v1` WebSocket instead of into a bastion's forward.
+
+### ◆ But the bridge is the relay's half, not the shared one
+
+Written into step 1 of this plan and wrong. Building it turned the sentence around: **the jump-host path
+needs no bridge from this repository at all.** `ForwardedPortDynamic` *is* the loopback listener — SSH.NET
+accepts on it, speaks SOCKS5 on it, and tunnels what it accepts through the bastion. Nothing is left for a
+bridge of ours to do. The relay is the case with no `SshClient` to hang a forward off, so it is the one that
+needs a listener written here, and it is the bridge's only consumer.
+
+What the two genuinely share is one step lower: **the connection being told to reach its target through a
+loopback SOCKS5 proxy, while staying about the target**. That is the piece both features stand on, it is
+fifteen lines in `BuildConnectionInfo`, and it is what step 1 turned out to be.
+
+So the bridge moves to step 3 and lands with the feature that uses it. Building it now would have been a
+component whose only caller was two steps away — which is the shape of the two defects this document is
+about.
+
+## The work, in order
+
+**0. Stop promising the relay.** ✅ The checkbox states that the relay is not wired up yet. One line on each
+head, and the only step that should ship on its own.
+
+**1. Dial through a loopback proxy.** ✅ `SshConnectionRequest` carries an optional `SshLoopbackProxy`, and
+`BuildConnectionInfo` builds SSH.NET's proxy `ConnectionInfo` when it is there. Three properties are worth
+knowing, and each is held by a test in `LoopbackProxyTests`:
+
+- **The type is a port and nothing else.** A proxy on any interface but loopback cannot be expressed, which
+  matters because the failure mode is an open SOCKS proxy on the user's network for the life of a shell,
+  and nothing would report it.
+- **SOCKS5 rather than a dumb pipe, so the target stays the target.** The host and port in the request are
+  the ones SSH.NET dials *through* the proxy and the ones the host key gate pins — so the same machine
+  reached through a bastion is pinned under its own name, not under `127.0.0.1:<ephemeral>`, which is not
+  an identity at all.
+- **A proxy that is not listening fails as a connection error**, not as an unknown host key. The gate
+  translates "no host key seen" into a fingerprint prompt, and a connection that never reached a server has
+  seen none either — so the prompt would offer to fix the wrong thing, with no fingerprint to show.
+
+Nothing calls it with a proxy yet. That is deliberate and it is one step wide: step 2 is the caller.
+
+**2. Jump hosts.** No server change, and no bridge. In order:
+
+- The hop's own connection, and a `ForwardedPortDynamic("127.0.0.1", 0)` started on it, whose `BoundPort`
+  becomes the `SshLoopbackProxy` for the connection after it. A chain of two is that twice.
+- `VaultViewModel` resolves `JumpHostIds` to hosts in the same vault, applying group inheritance per hop the
+  way the target already gets it, and refuses a chain that crosses a vault — the same refusal the group
+  picker already makes, for the same reason.
+- ◆ **The target's address is resolved at the last hop, not here.** A SOCKS CONNECT names the target and the
+  bastion resolves it, so what has to be stored on the host is the address *the bastion* can reach — which
+  is what an `ssh_config` means by `HostName` beside a `ProxyJump`, and what the importer is already
+  carrying across verbatim. Nothing needs to change for that to be true; it needs to be said, because a
+  host that resolves here and not there fails as a SOCKS "general failure" naming neither end.
+- Per-hop host keys. Each hop is a separate handshake against a separate endpoint, so the pin, the unknown
+  key prompt and the changed-key refusal run per hop. **The prompt has to name which hop it is about**, or
+  somebody approves a bastion's fingerprint believing it is the target's — see `HostKeyCard`, which is built
+  around one connection and one question.
+- Per-hop credentials, including a hop that wants a typed password. `IsAskingForConnectPassword` asks about
+  one host today.
+- Teardown: the hops belong to the outer session and go with it, including when the outer connect fails
+  half way. A leaked bastion connection is an open session on a machine the user believes they left.
+- The schema version. A chain becomes a real field, so it joins the ladder in `HostSecretCodec` — a host
+  carrying one must not be editable by a client that would drop it. That is the whole point of the rule.
+- The editor: a picker over other hosts in the same vault, and the host detail's subtitle finally getting
+  the `⤷ bastion-eu` the design asked for.
+
+**3. The relay, and the bridge with it.** A loopback `TcpListener` on an ephemeral port that accepts one
+connection, answers a SOCKS5 CONNECT on it, and pipes the rest into the `dodossh.relay.v1` WebSocket —
+SOCKS5 rather than a raw pipe so that this path presents the same interface step 1 already speaks, and the
+target's identity stays the target's. Then `POST /relay/tickets` with the host id and the WebSocket with the
+ticket in `Sec-WebSocket-Protocol`; the ticket is single-use and expires in 30 seconds, so it is fetched per
+connect and never cached. The bridge binds `127.0.0.1` explicitly, accepts once and stops listening. Then
+the sentence step 0 added comes out of both heads.
+
+**4. File transfer.** `ISftpSessionFactory.OpenSftpAsync` opens its own second connection, so a host that
+needs a chain or a relay to reach needs it there too, or SFTP silently fails for exactly the hosts this
+work exists for. ✅ for the relay.
+
+## What step 3 built
+
+**The server**, in `DodoSSH.Api/Features/Relay`:
+
+- `POST /api/v1/relay/tickets` makes every decision. It finds the host, requires Connect on its vault and
+  the relay switched on, resolves the stored name, and runs every answer through `RelayAddressPolicy` —
+  the non-overridable list first, then the operator's denials, allow-list and private-network switch. What
+  passes goes into a Data Protection ticket; the name is not resolved again. **Every target refusal is one
+  answer**, 403 `relay-target-rejected` with one sentence, because anybody who can write a host can point
+  it anywhere, and a distinct "resolves, but refused" would be a scanner for the operator's DNS. The real
+  reason is logged (event 2301).
+- `GET /api/v1/relay/connect` is anonymous at the HTTP layer and takes the ticket as the second
+  subprotocol. It verifies, checks expiry against `TimeProvider`, and spends the ticket with a unique
+  insert into `relay_ticket_use` before accepting the upgrade — so a refused ticket is a 401 and no socket.
+  After the upgrade, a refusal the client could not have avoided is a close code: 4429 for a limit, 4502
+  for a host the server could not reach, 4408 idle, 4409 lifetime, 1001 at the end of a shutdown's drain.
+- `relay_session` is the audit row: opened before the dial, closed after, each from a scope of its own.
+
+**The client:**
+
+- `RelayBridge` (Client.Ssh) is the loopback SOCKS5 listener. It binds `127.0.0.1`, accepts once, answers
+  one CONNECT without acting on its target, and pipes to the WebSocket. It is handed to the connection as
+  the new `SshLoopbackProxy.Lifetime`, so the connection disposes it on every way a session ends, and a
+  failed connect disposes it before throwing.
+- `ServerRelay` (Client.Session) asks for the ticket, opens the socket and starts the bridge, and turns
+  every refusal into a sentence. It is `IVaultServer.Relay`, decided from `/api/v1/meta` the way `Events`
+  is; a server without the feature supplies `UnavailableServerRelay`.
+- `VaultViewModel` opens a relay for a host with the box ticked and **never falls back to dialling it
+  directly** — the user ticked it because that path does not work, or goes somewhere they do not mean. A
+  relay that closed with a reason replaces SSH.NET's transport error with that reason, and the
+  reachability probe is skipped for a relayed host, because it would test the direct path.
+
+### ◆ Two things that are not what ADR 0004 wrote
+
+- **No `System.IO.Pipelines`.** Each direction reads one buffer and does not read again until the write
+  has completed. That is the same end-to-end backpressure — a slow peer stops the read, which fills the
+  kernel's window, which throttles the far sender — with a tighter bound: two 32 KiB buffers a session
+  rather than ~2 MiB.
+- **No `?ticket=` fallback, and no device id in the ticket.** The fallback exists for clients that cannot
+  set a subprotocol, which neither head is, and it would put a bearer credential in the one place — the
+  URL — that access logs keep. A device is never visible to the server on an ordinary request.
+
+### ◆ A WebSocket operation is never cancelled, on either side
+
+Cancelling a pending receive or send on a .NET WebSocket aborts it, and an aborted socket sends no close
+frame. The close frame is the only way the client learns that the server could not reach the host, so both
+the server's pump and the client's bridge stop a direction by closing and waiting a bounded time instead.
+`AnIdleSession_IsClosedWithTheIdleCode` holds the server to it: it asserts the code arrives, not merely that
+the socket ended.
+
+### What a developer meets
+
+The relay refuses loopback with no configuration escape, so the development stack's sshd on
+`localhost:2222` cannot be reached through it. Point a relayed host at an address the API process reaches
+on a real interface — the machine's own LAN address works — or at the container's address on the Docker
+network.
+
+## Traps already known
+
+**A relay host may not inherit its port, and a jump host has no such rule.** `TryValidate` enforces the
+first because the server stores the port and a group edit would silently change what the relay dials. The
+chain has no plaintext counterpart, so it inherits normally — do not copy the restriction across out of
+symmetry.
+
+**Two hosts can name each other.** A chain is ids, and nothing stops A jumping through B while B jumps
+through A. Resolve iteratively with a visited set and refuse a cycle before dialling anything, rather than
+discovering it as a stack overflow inside a connect.
+
+**The bastion's own group defaults matter.** A hop is a host, so it resolves its port, username and binding
+through `HostInheritance` exactly as the target does. Skipping that dials 22 as nobody on a bastion that is
+on 2222 as `deploy`.
+
+**`ForwardedPortDynamic(0)` reports its port, and the one-argument constructor's bind address is
+undocumented.** The first is now pinned by a test rather than trusted. The second is why the two-argument
+constructor is always used: if the default were `0.0.0.0` the failure would not be a test failure, it would
+be a SOCKS proxy into the developer's network that nothing reports, so the bound host is asserted too.
+
+**◆ The test server refuses forwarding, and says so nowhere useful.** `linuxserver/openssh-server` ships
+`AllowTcpForwarding no`. A dynamic forward starts anyway — opening one asks the server nothing — and every
+connection through it is then refused at channel-open, which SSH.NET surfaces as `SOCKS5: General failure`,
+naming neither the server nor the setting. `SshServerFixture` patches it after start and HUPs sshd. **There
+are two `sshd_config` files in that image** and the running server uses `/config/sshd/sshd_config`; patching
+`/etc/ssh/sshd_config`, which is the one a search finds first, changes the text and nothing else.
+
+**The relay bridge and the jump bridge are the same class and not the same lifetime.** A ticket is
+single-use with a 30-second expiry; a bastion's forward lives as long as the session. Sharing the listener
+is right, sharing a lifetime policy is not.
+
+## Tests
+
+- ✅ A connect through a real SOCKS5 forward to a real `sshd`, in `LoopbackProxyTests`. **One container,
+  used as both ends** — the forward is opened on a connection to the fixture's server and the connection
+  under test goes back to the same server through it. A second container would look more like the topology
+  and establish nothing extra: what is under test is that the proxy is honoured, that the target is what
+  gets pinned, and that a failure on the way through is reported as itself.
+- A cycle in a chain is refused before any socket is opened.
+- Each hop's host key is asked about separately, and the question names the hop.
+- A chain crossing a vault is refused with a reason, as the group picker's is.
+- The bridge binds loopback only — assert the bound address, because the failure is silent and the
+  consequence is an open proxy.
+- SFTP to a host behind a chain, once step 4 lands.
+- Mutations that must fail something: bind the bridge on `IPAddress.Any`; drop the visited set; skip group
+  inheritance for a hop; and tear down the outer session without the hops.
+
+## Prose that becomes false
+
+- `docs/design-import-gaps.md` — the host subtitle's `⤷ bastion-eu` row, the SFTP `sftp over bastion-eu`
+  row, and the status bar's `via bastion-eu` row, all of which say jump hosts are data-only.
+- `Client.Import/ImportedHost.cs` — the note written onto every imported host with a `ProxyJump`, and the
+  remark above it.
+- `README.md` and `docs/android-port.md` wherever the relay is described as available. ✅ at step 3.
+- ADR 0004 gains a note that its last consequence was built, and how. ✅ at step 3.
+
+## What the first draft of this document got wrong
+
+It recommended deleting `JumpHostIds`, on the evidence that nothing writes it, nothing reads it, and it is
+missing from the schema-version ladder. The first two facts are true and the conclusion did not follow.
+
+Two things were missed. **ADR 0004 had already designed the implementation** — the loopback bridge, the
+SOCKS5 dynamic forward, "one mechanism, two features" — so the transport was a solved problem sitting in an
+accepted ADR, and the fortnight that draft estimated was priced without it. And **the stored shape is
+right**: an ordered list of host ids is exactly what a chain is, the merge arm is already correct, and the
+missing schema version is a line to add rather than evidence of a bad model.
+
+The lesson is narrower than "read the ADRs": it is that *nothing reads this field* was taken as evidence the
+field was a mistake, when it was evidence of an unfinished feature — and the same reasoning applied one
+paragraph further would have found the relay checkbox, which is the same shape and is actively lying to
+users.
+
+**And then step 1 was wrong too**, in the same direction: it named the bridge as the shared foundation on
+the strength of ADR 0004's "one mechanism, two features", without checking which half of the mechanism
+SSH.NET already provides. It provides the jump host's half entirely. The shared piece was one level down and
+a tenth of the size, and the bridge belongs with the relay. Both corrections came from writing the code
+rather than from reading more — which is an argument for the step-at-a-time ordering rather than against
+planning, but only if the plan is edited when a step answers back.

docs/unlocking-without-the-passphrase.md

@@ -0,0 +1,195 @@
+# Unlocking without the passphrase
+
+Every account here is issued a recovery code at enrollment. It is generated on the client, it wraps the
+identity bundle, the server stores that wrap, and both heads go to some trouble to make sure the user writes
+it down — the phone raises `FLAG_SECURE` for that screen alone and refuses to let anybody click past it.
+
+**Nothing can use it.** There is no code path in this product that opens a recovery wrap. This document is
+the plan for the change that fixes that, and it is written to be picked up cold.
+
+> **Status: planned. None of the six steps below is built.**
+>
+> | Step | State | Notes |
+> | --- | --- | --- |
+> | 1. The endpoint that serves the wrap | Not started | `GET /api/v1/me/recovery-wrap`, and deliberately not `/me` |
+> | 2. `SessionOpener.UnlockWithRecoveryAsync` | Not started | The unwrap, then the path `UnlockAsync` already takes |
+> | 3. Setting a new passphrase | Not started | `PUT /api/v1/me/wrap`. Delivers *change passphrase* as well |
+> | 4. Both heads | Not started | A way in from the unlock screen, and a box for the code |
+> | 5. The prose that becomes false | Not started | Three shipped claims disagree with each other today |
+> | 6. `crypto.md`'s status note | Not started | "One of four ways" is one of two, and will be one of three |
+
+## What is wrong
+
+Walk the failure through, because it is worse than a missing feature.
+
+Forget the passphrase and the identity bundle cannot be unwrapped. No bundle means no vault keys, and no
+vault keys means every item in every vault is unreadable. Signing out and back in does not help: the server
+hands back the same passphrase wrap. The device key would be the other door, and **sign-out withdraws it** —
+which is the advice the unlock screen gives for exactly this situation. The recovery wrap sitting on the
+server is the only thing left, and nothing opens it.
+
+So the loss is total and permanent, and the thing built to prevent it is inert.
+
+Meanwhile the product says three things about this, and they do not agree with one another:
+
+| What it says | Where | True today |
+| --- | --- | --- |
+| "The code is the only thing standing between a forgotten passphrase and an unrecoverable vault" | `docs/manual-checks.md` §10.2 | **No.** It stands between nothing |
+| "losing it *along with* the passphrase means the vault is unrecoverable" | `docs/android-port.md`, state 4 | **No.** It implies the code alone is enough |
+| "Signing out … is the only answer to a forgotten passphrase — nothing can recover one" | `README.md`, Locking | Yes, and it contradicts both of the above |
+
+The third is the honest one. That is the state this plan changes, and until it does, the first two are the
+two sentences in this repository most likely to cost somebody their keychain.
+
+## What already exists
+
+More than half of it, which is why this is worth doing now rather than treating as a feature.
+
+- **The code.** `ClientEnrollment.GenerateRecoveryCode` — 160 bits, base32 over a 32-character alphabet with
+  `I`, `L`, `O` and `U` left out so a transcription cannot land on a different valid code, printed as
+  32 characters in groups of five.
+- **The wrap.** The same class derives `KEK_rc` under `Argon2Profile.RandomSecret` (64 MiB, 3 passes) and
+  sends `RecoveryWrappedPrivateKey` and `RecoveryKdfParameters` with the enrollment.
+- **The row.** `EnrollmentService` stores it as `UserKeyWrapKind.Recovery`, beside the passphrase and device
+  wraps.
+- **The specification.** `docs/crypto.md` §2 gives the parameters for `KEK_rc` and §3 puts the wrap in the
+  key hierarchy beside the passphrase and device ones. Nothing below needs a spec change.
+- **The cache.** `LocalCacheKey` derives from the identity bundle under `dsh1/localcache/v2`, not from `MK`
+  — see the changed-2026-07-30 note in `crypto.md` §3.2. That change was made partly *for* this: a recovery
+  unlock derives a different `MK` and would otherwise open the identity and then fail to read the cache it
+  had itself written. It is already paid for and currently untested from this direction.
+
+## What is missing, exactly
+
+| | Gap | Where it lands |
+| --- | --- | --- |
+| A | Nothing serves the wrap. `MeResponse` carries `WrappedPrivateKey` and `KdfParameters` — the passphrase pair, and only that | `DodoSSH.Contracts`, `Api/Features/Identity` |
+| B | No unlock path. `SessionOpener` has `UnlockAsync` and `UnlockWithDeviceAsync`, and nothing else | `Client.Session` |
+| C | No way to set a new passphrase afterwards. No endpoint, no client path, no UI | server + client |
+| D | No way in from either unlock screen | `UnlockCard.axaml`, Android `LockedScreen.axaml` |
+
+The device wrap is not a precedent for A: it is cached locally in `StoredUnlockMaterial` at the moment it is
+registered, so it never has to be fetched.
+
+## The decisions, and the reasons
+
+**A separate endpoint, not `/me`.** `GET /api/v1/me/recovery-wrap`, called only when somebody has said they
+have forgotten their passphrase. `/me` is fetched by every client at every launch, and putting the recovery
+wrap in it would hand that blob to anybody holding a stolen OIDC session, permanently, for no benefit. The
+wrap's real defence is Argon2id over 160 bits and it does not stop being safe in the response — this is the
+cheaper rule of not serving what nothing needs. It answers 404 for an account with no recovery row, which is
+every account enrolled by a client that did not send one.
+
+**Online only, and the screen says so.** The wrap is deliberately *not* added to `StoredUnlockMaterial`.
+Caching it would put a second door on every laptop, protected by a weaker KDF profile than the passphrase
+one, for a case that already requires a network — recovery begins with signing in.
+
+**Setting a new passphrase is part of the flow, not a follow-up.** Without it the account unlocks with a
+one-time code forever, and the code is a thing people keep on paper. This is the same re-wrap a *change
+passphrase* feature needs, so step 3 delivers both; change-passphrase is then a button, not a project.
+
+**The identity key is not rotated.** Rotating it would invalidate every vault key grant the account holds
+and force a rekey of every vault it can read — see [ADR 0010](adr/0010-vault-key-rotation.md) for what one
+of those costs, and it is per vault. Nothing about the bundle is compromised by its owner proving possession
+of it, so there is nothing to rotate away from. Identity key rotation is M5 and stays there.
+
+**The recovery code is not re-issued after use.** Issuing a new one means asking somebody to write down a
+new code at the exact moment they have demonstrated they lose things, and the old code is not weakened by
+having been typed. "Issue a new recovery code" is a separate feature with its own screen, and it is the
+right place for that question.
+
+**Rate limiting is wanted here and is not a blocker.** This is the first endpoint in the product with a
+guessable secret behind it, and `docs/platform-flags.md` records that rate limiting is unimplemented (M2).
+Argon2id at 64 MiB is the cost that matters — a guessing attack pays it per attempt — so this ships without
+and the endpoint is the reason to do the M2 item next.
+
+## The work, in order
+
+Each step compiles with the whole suite green before the next begins.
+
+1. **The endpoint.** `GET /api/v1/me/recovery-wrap`, returning the wrap and its `KdfParameters`, or 404. A
+   contract type, an entry in `PublicAPI.Unshipped.txt`, and a `SyncEndpointTests`-style refusal test for an
+   account without one.
+2. **The unlock.** `SessionOpener.UnlockWithRecoveryAsync(string code)`: canonicalise the code, derive
+   `KEK_rc` from the served parameters, unwrap the bundle, and then join the path `UnlockAsync` already
+   takes from the moment it holds one. Nothing after the unwrap should be new code — if it is, the two
+   unlocks have diverged and one of them is wrong.
+3. **The re-wrap.** `PUT /api/v1/me/wrap` taking a new passphrase wrap and its parameters, and the client
+   half that derives the new `KEK_pp` and calls it. Forced as the last step of a recovery: the session is
+   open, so refusing to go further until a passphrase is set costs nothing and is the only moment the user
+   is certainly present.
+4. **Both heads.** A "forgotten your passphrase?" route from `UnlockCard` and from Android's
+   `LockedScreen`, a box that accepts the code as printed, and the new-passphrase form after it. The phone's
+   version has to survive the keyboard covering the box — `docs/manual-checks.md` §10.3 lists the five boxes
+   that already do, and this is a sixth.
+5. **The prose.** Below.
+6. **`crypto.md`.** A status note, not a spec change — see the last section.
+
+## Traps already known
+
+**◆ The code is derived from the string exactly as displayed, dashes included.** `GenerateRecoveryCode`
+says so where it builds the groups: the separators are not decoration to be stripped before hashing, they
+are part of the input. So step 2 must *canonicalise to the printed form* rather than normalise it away —
+up-case, then re-group in fives — and it must not simply strip dashes and hash what is left. Getting this
+wrong produces a code that verifies nowhere and an error message that says the code is wrong.
+
+**The alphabet has holes in it.** `I`, `L`, `O` and `U` are not in it. A user who writes `O` for `0` should
+be met with a code that works, so the canonicaliser should fold the four missing letters onto their
+look-alikes before deriving. That is a deliberate leniency and belongs beside the alphabet's own comment.
+
+**The two KDF profiles are not the same, and the stored parameters are what to use.** The passphrase wrap is
+`PassphraseDefault` (256 MiB, 4 passes) and the recovery wrap is `RandomSecret` (64 MiB, 3), which is
+correct — 160 random bits do not need the same stretching as a chosen phrase. Derive from the parameters the
+server served with the wrap, never from a profile constant, or an account enrolled by a client with
+different settings will refuse a valid code.
+
+**The cache property is untested from this direction.** A recovery unlock derives a different `MK` and must
+still read a cache written under a passphrase unlock. That works today because `LocalCacheKey` hangs off the
+bundle, and no test asserts it, because nothing could reach the state. Assert it in step 2 rather than
+trusting the note in `crypto.md`.
+
+**Sign-out is the neighbouring path and must stay reachable.** The unlock screen offers signing out as the
+answer to a forgotten passphrase. It stays: recovery needs the code, and somebody without it still needs a
+way to hand a laptop on. What changes is that it stops being the *only* answer, so the copy on both screens
+has to place them beside each other rather than replacing one with the other.
+
+## Tests
+
+- Round trip: enrol, recover, unlock, set a new passphrase, unlock again with it.
+- The old passphrase stops working after step 3, and the recovery code still does.
+- A wrong code is refused without disclosing whether the account has a recovery wrap at all.
+- The cache written under a passphrase unlock is readable after a recovery unlock — the `localcache/v2`
+  property, asserted rather than assumed.
+- An account with no recovery row: the endpoint 404s and the screen says the code was never issued rather
+  than that it is wrong.
+- One `SystemTests` pass against the real API, because this is a two-endpoint flow and that suite is the one
+  that consumes the artefacts that ship.
+- Mutations to confirm each test can fail: reverse the canonicalisation so dashes are stripped; derive from
+  `Argon2Profile.RandomSecret` instead of from the served parameters; and skip the forced re-wrap in step 3.
+
+## Prose that becomes false
+
+Not a tidy-up. Each of these currently tells a user something that this change makes wrong in the other
+direction, and two of them are wrong right now.
+
+- `docs/manual-checks.md` §10.2 — "the only thing standing between a forgotten passphrase and an
+  unrecoverable vault". True once this ships, false today. It also needs a new check: typing a code from
+  paper, which is the one part of this no test can perform.
+- `docs/android-port.md`, state 4 — "losing it along with the passphrase means the vault is unrecoverable".
+- `README.md`, Locking — "Signing out … is the only answer to a forgotten passphrase — nothing can recover
+  one." Correct today and the sentence this change exists to falsify.
+- `docs/design-import-gaps.md`, Preferences — "a forgotten passphrase — which is unrecoverable by
+  construction, so the unlock screen would otherwise be a dead end", which is the stated justification for
+  signing out being on that screen. The justification survives; the parenthesis does not.
+- Both unlock screens, wherever they name signing out as the answer.
+
+## The spec's "four ways"
+
+`docs/crypto.md` §3.2 says a passphrase is "only one of four ways to open a vault" — passphrase, device,
+recovery and escrow. Two exist. This change makes it three; escrow is M5 by decision and has an enum member
+and nothing else.
+
+**Do not edit that sentence.** It is normative and it is about the DSH1 format, where four ways is exactly
+right. What is missing is a statement about *this build*, so add a short table under §3.2 saying which wrap
+kinds this client can create and which it can open, and update it here rather than in the frozen prose. Land
+it in the same commit as step 2, so the two can never disagree again.

scripts/ci-linux-rpm.sh

@@ -0,0 +1,385 @@
+#!/usr/bin/env bash
+#
+# Turns a published linux-x64 tree into an RPM, and then checks the RPM rather than the exit code.
+#
+# Called from two jobs in .github/workflows/ci.yml and written once for the reason the four copies of
+# "ensure node and git" in that file are not: those cannot be factored out — a composite action would
+# need the checkout it exists to unblock — and this can. The build job packages the release-shaped
+# RPM and throws it away, exactly as it does for Windows and macOS; the desktop job packages the
+# nightly-shaped one and publishes it. Everything that differs between those two is an argument.
+#
+# It is also runnable at a desk, which is the point of it being a script:
+#
+#     dotnet publish src/DodoSSH.Client.App/DodoSSH.Client.App.csproj \
+#       --configuration Release --runtime linux-x64 --self-contained true \
+#       -p:RestoreLockedMode=false --output /tmp/linux-x64
+#     scripts/ci-linux-rpm.sh --payload /tmp/linux-x64 --version 1.2.3 --out /tmp/rpm
+#
+# Nothing here holds a secret and nothing here uploads anything. Publishing is the calling job's, and
+# ADR 0013 rule 3 is what decides which channel may do it.
+
+set -euo pipefail
+
+repo_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+
+payload=""
+version=""
+channel="release"
+out=""
+
+while [ $# -gt 0 ]; do
+    case "$1" in
+        --payload) payload="$2"; shift 2 ;;
+        --version) version="$2"; shift 2 ;;
+        --channel) channel="$2"; shift 2 ;;
+        --out)     out="$2";     shift 2 ;;
+        *) echo "Unknown argument: $1" >&2; exit 2 ;;
+    esac
+done
+
+for required in payload version out; do
+    if [ -z "${!required}" ]; then
+        echo "--$required is required." >&2
+        exit 2
+    fi
+done
+
+case "$channel" in
+    release|nightly) ;;
+    *) echo "--channel must be release or nightly, not '$channel'." >&2; exit 2 ;;
+esac
+
+# The apphost, named rather than globbed. A publish that produced no apphost at all would otherwise
+# package happily into an RPM whose /usr/bin symlink points at nothing, and the first thing anybody
+# would learn about it is a shell saying "No such file or directory" for a file that is right there.
+if [ ! -s "$payload/DodoSSH" ]; then
+    echo "There is no apphost at $payload/DodoSSH, so there is nothing to package." >&2
+    ls -la "$payload" >&2 || true
+    exit 1
+fi
+
+# ◆ THE FOUR THINGS THAT DIFFER BETWEEN THE CHANNELS, AND WHY IT IS FOUR AND NOT ONE.
+#
+# This is ADR 0013 rule 9's table, in the spelling this platform uses. The package name keeps dnf from
+# treating one as an upgrade of the other; the directory under /opt keeps the two trees apart on disk;
+# the /usr/bin name and the desktop entry keep them apart in front of a person, who would otherwise
+# have two identical launchers and no way to tell which was which.
+#
+# The fifth separation — the profile directory holding the vault cache, the outbox and the device key
+# — is not here and must not be, because it belongs to the build rather than to the package: the
+# nightly is published with DodoChannel=nightly, and ClientPaths reads that to suffix the directory.
+# Doing it here as well would be a second answer to a question that already has one.
+if [ "$channel" = nightly ]; then
+    name=dodossh-nightly
+    dir=DodoSSH.Nightly
+    bin=dodossh-nightly
+    title="DodoSSH Nightly"
+else
+    name=dodossh
+    dir=DodoSSH
+    bin=dodossh
+    title="DodoSSH"
+fi
+
+# ◆ SEMVER IS NOT AN RPM VERSION, AND THE TRANSLATION IS THE ONE PLACE THIS CAN GO QUIETLY WRONG.
+#
+# rpm forbids `-` in a version, which is exactly the character SemVer uses to introduce a prerelease.
+# MinVer's answer on main is 0.0.1-alpha.0.143, so the naive move — strip the suffix — would give
+# every nightly the version 0.0.1 and make each one look like a reinstall of the last.
+#
+# `~` is rpm's own answer and has been since 4.10: it sorts *before* the empty string, so
+# 0.0.1~alpha.0.143 is older than 0.0.1 the way alpha.143 is older than the release. Heights keep
+# rising within a floored version and the first real tag moves past all of them, which is the same
+# ordering property the Velopack channels have and is worth having on both.
+#
+# Build metadata after `+` is dropped rather than translated. It is the one part of a SemVer string
+# that is defined as not participating in ordering, so keeping it would put a character rpm has no
+# rule for into the field ordering is done on.
+rpm_version="${version%%+*}"
+rpm_version="${rpm_version/-/\~}"
+
+case "$rpm_version" in
+    *-*)
+        echo "The version $version translates to $rpm_version, which still contains a hyphen." >&2
+        echo "rpm cannot order a version with one in it; this needs a rule rather than a retry." >&2
+        exit 1
+        ;;
+esac
+
+# Always 1. The release field distinguishes several packagings of one upstream version, and there is
+# only ever one packaging here — the version already carries the build height, because upstream and
+# packager are the same repository.
+rpm_release=1
+
+echo "==> $name $rpm_version-$rpm_release.x86_64, from $payload"
+
+# ◆ rpmbuild ON A MUSL HOST, WHICH IS FINE, AND WHY IT IS WORTH SAYING SO.
+#
+# The runner is Alpine. Nothing about that shows up in the package: rpmbuild composes an archive and
+# some metadata, and the one part of it that reads the payload — elfdeps, which the spec leaves on for
+# requires — parses ELF headers with libelf rather than loading anything, so it reads x86_64 glibc
+# binaries on a musl host correctly. Building in a Fedora container was the alternative and would
+# have meant `docker cp`-ing a 130 MB publish tree in and the package back out, because this runner is
+# itself a container and a bind mount resolves on the daemon's host rather than here — see the note on
+# the android job for the hour that cost.
+if ! command -v rpmbuild >/dev/null 2>&1; then
+    SUDO=""
+    [ "$(id -u)" -eq 0 ] || SUDO="sudo"
+
+    echo "==> installing rpm"
+    if command -v apk >/dev/null 2>&1; then
+        $SUDO apk add --no-cache rpm
+    elif command -v dnf >/dev/null 2>&1; then
+        $SUDO dnf install -y rpm-build
+    elif command -v apt-get >/dev/null 2>&1; then
+        $SUDO apt-get update -qq
+        $SUDO apt-get install -y --no-install-recommends rpm
+    else
+        echo "No apk, dnf or apt-get here, so rpmbuild cannot be installed from inside the job." >&2
+        echo "Add rpm to the runner's image; the Linux package cannot be built without it." >&2
+        exit 1
+    fi
+fi
+rpmbuild --version
+
+# ◆ THE LIST OF PRIVATE SONAMES, WORKED OUT FROM THE TREE RATHER THAN WRITTEN DOWN.
+#
+# build/linux/dodossh.spec turns rpm's automatic provides off for everything under /opt, because the
+# libraries there are this application's private copies and have no business satisfying anybody
+# else's dependency. The requires side has to be told the same thing about the same set, or a bundled
+# library that links another bundled library asks for a soname nothing provides and the package
+# becomes uninstallable on every host — which .NET's own libmscordbi.so, linking libmscordaccore.so,
+# does on every publish.
+#
+# Derived here because the set belongs to the payload. A list in the spec would be Microsoft's and
+# SkiaSharp's internal file names, maintained by hand, going stale on the first runtime bump — and
+# going stale in the direction that produces a package nobody can install.
+#
+# Only the top level is scanned, which is where a .NET publish puts its native libraries. The dots
+# are escaped because this becomes a regular expression; nothing else in a soname is a metacharacter.
+#
+# A shell glob rather than `find -printf`, which is GNU find's and not busybox's — the runner is
+# Alpine, where that flag is a usage message and an empty list, which would silently disable the whole
+# exclusion and put the uninstallable package back.
+bundled="$(
+    for library in "$payload"/lib*.so*; do
+        [ -f "$library" ] && basename "$library"
+    done |
+        sed 's/\./\\./g' |
+        sort -u |
+        paste -sd '|' -
+)"
+
+if [ -z "$bundled" ]; then
+    echo "There are no native libraries at the top of $payload." >&2
+    echo "A self-contained .NET publish always has some, so this is a publish that went wrong" >&2
+    echo "rather than a package with nothing to exclude." >&2
+    exit 1
+fi
+
+work="$(mktemp -d)"
+trap 'rm -rf "$work"' EXIT
+
+mkdir -p "$out"
+
+# --target rather than trusting the host's arch, because the host's arch here is whatever the runner
+# happens to be and the payload's is decided by the RID the publish asked for. They agree today and
+# an aarch64 runner would silently make them disagree.
+#
+# `dist` is emptied deliberately. It is the %{?dist} tag that puts .fc42 in a filename, and it names
+# the distribution the package was *built* on — which for a self-contained tree built on Alpine would
+# be both meaningless and, if rpm guessed something plausible, actively misleading.
+rpmbuild -bb "$repo_root/build/linux/dodossh.spec" \
+    --target x86_64 \
+    --define "_topdir $work" \
+    --define "_rpmdir $out" \
+    --define "dist %{nil}" \
+    --define "dodo_name $name" \
+    --define "dodo_version $rpm_version" \
+    --define "dodo_release $rpm_release" \
+    --define "dodo_dir $dir" \
+    --define "dodo_bin $bin" \
+    --define "dodo_title $title" \
+    --define "dodo_payload $payload" \
+    --define "dodo_bundled $bundled" \
+    --define "dodo_icons $repo_root/src/DodoSSH.Client.App/Assets" \
+    --define "dodo_udev $repo_root/build/linux/71-dodossh-tpm.rules" \
+    --define "dodo_licence $repo_root/LICENSE"
+
+# rpmbuild writes into <_rpmdir>/<arch>/, and every caller wants the file rather than the tree.
+rpm_file="$out/x86_64/$name-$rpm_version-$rpm_release.x86_64.rpm"
+if [ ! -s "$rpm_file" ]; then
+    echo "rpmbuild reported success and there is no package at $rpm_file." >&2
+    find "$out" >&2 || true
+    exit 1
+fi
+mv "$rpm_file" "$out/"
+rmdir "$out/x86_64" 2>/dev/null || true
+rpm_file="$out/$(basename "$rpm_file")"
+
+echo "==> $(basename "$rpm_file"), $(wc -c < "$rpm_file") bytes"
+
+# ◆ AND NOW THE PART THAT IS NOT PACKAGING: ASKING THE PACKAGE WHAT IT SAYS ABOUT ITSELF.
+#
+# Everything above can succeed and still produce something nobody can install, and none of the ways
+# that happens are visible in an exit code. Each check below is one of them.
+#
+# ◆ EVERY READER IN A PIPELINE HERE MUST CONSUME ALL OF ITS INPUT, AND THAT IS NOT A STYLE RULE.
+#
+# `set -o pipefail` at the top of this file means a pipeline's status is the rightmost non-zero one.
+# A reader that stops early — `grep -q`, `head -n` — closes the pipe while the writer is still
+# writing, the writer dies of SIGPIPE with status 141, and pipefail hands that 141 to the `if`. So a
+# check that found what it was looking for reports the opposite, and a diagnostic printed on the way
+# to `exit 1` kills the script under `set -e` before it gets there.
+#
+# It is a threshold rather than a constant, which is what made it look like somebody else's bug: the
+# writer only blocks once it has more than a pipe buffer left to write, about 64 KB on Linux. `$files`
+# is one line per packaged file, so this passed for as long as the published tree was small and then
+# failed on the commit that pushed the path text past 64 KB — reporting `/opt/DodoSSH/DodoSSH` missing
+# from a package that plainly contained it, and exiting 141 rather than the 1 it meant to.
+#
+# Hence `grep pattern >/dev/null` rather than `grep -q pattern`, which has the same exit status and
+# reads to the end; and `awk 'NR<=40'` rather than `head -40`, which prints the same lines and reads
+# to the end. Nothing here is allowed to stop early, and `|| true` is not the fix — it would hide a
+# real failure of the writer alongside the imaginary one.
+
+requires="$(rpm -qp --requires "$rpm_file")"
+provides="$(rpm -qp --provides "$rpm_file")"
+files="$(rpm -qp --list "$rpm_file")"
+
+# 1. The bundled-soname trap the spec's provides exclusion sets. If some library under /opt ever
+#    gains a DT_NEEDED on another library under /opt, elfdeps will put that soname in the requires
+#    while the exclusion keeps it out of the provides, and the result is a package no host on earth
+#    can satisfy. Nothing about it looks wrong until an install fails, so it is checked here against
+#    the actual contents rather than assumed.
+carried="$(printf '%s\n' "$files" | sed -n 's|^/opt/.*/\(lib[^/]*\.so[^/]*\)$|\1|p' | sort -u)"
+unsatisfiable=""
+for soname in $carried; do
+    if printf '%s\n' "$requires" | grep -F "$soname(" >/dev/null; then
+        unsatisfiable="$unsatisfiable $soname"
+    fi
+done
+if [ -n "$unsatisfiable" ]; then
+    echo "This package requires sonames it also carries and does not provide:$unsatisfiable" >&2
+    echo "That is unsatisfiable on every host, and it means the requires exclusion computed above" >&2
+    echo "no longer matches what rpm generated — a bundled library outside the top level, or an rpm" >&2
+    echo "that has stopped honouring __requires_exclude. See build/linux/dodossh.spec." >&2
+    exit 1
+fi
+
+# 2. The provides exclusion still doing its job. A package advertising libcoreclr.so to the whole
+#    system is the thing that exclusion exists to prevent, and a future rpm that stopped honouring
+#    the macro would break it silently and in the direction nobody looks.
+if printf '%s\n' "$provides" | grep -E '^lib.*\.so' >/dev/null; then
+    echo "This package advertises bundled libraries to the system:" >&2
+    printf '%s\n' "$provides" | grep -E '^lib.*\.so' >&2
+    echo "__provides_exclude_from in build/linux/dodossh.spec is meant to stop that." >&2
+    exit 1
+fi
+
+# 3. elfdeps having actually run. If it silently did nothing — a runner image whose rpm was built
+#    without libelf would do exactly that — the package would carry no glibc requirement at all and
+#    would install cheerfully onto a host too old to run it. libc.so.6 is the one dependency a
+#    self-contained .NET tree cannot fail to have.
+if ! printf '%s\n' "$requires" | grep '^libc\.so\.6' >/dev/null; then
+    echo "This package does not require libc, which a self-contained .NET tree cannot be true of." >&2
+    echo "The automatic ELF scan produced nothing; rpm here is probably built without elfdeps." >&2
+    printf '%s\n' "$requires" >&2
+    exit 1
+fi
+
+# 4. The four files a person interacts with, named individually. The tree under /opt is thousands of
+#    entries and its absence would be caught by size alone; these are the small ones whose absence
+#    would leave a package that installs, occupies 130 MB and puts nothing on the menu.
+for expected in \
+    "/opt/$dir/DodoSSH" \
+    "/usr/bin/$bin" \
+    "/usr/share/applications/$bin.desktop" \
+    "/usr/share/icons/hicolor/256x256/apps/$bin.png" \
+    "/usr/lib/udev/rules.d/71-dodossh-tpm.rules"
+do
+    if ! printf '%s\n' "$files" | grep -xF "$expected" >/dev/null; then
+        echo "The package was built and $expected is not in it." >&2
+        printf '%s\n' "$files" | awk 'NR<=40' >&2
+        exit 1
+    fi
+done
+
+# 5. Nothing world-writable, which on this package is a security property rather than a lint rule.
+#    The spec sets every mode explicitly and says why; this is the check that it worked, because the
+#    failure it guards against arrives by inheritance from a publish directory rather than by anybody
+#    changing this. A world-writable /opt/DodoSSH lets any local account replace the SSH client every
+#    other account on the machine runs.
+#
+#    Symlinks are excluded because they are always 777 on Linux and the mode means nothing; what the
+#    link points at is what has permissions.
+writable="$(
+    rpm -qp --queryformat '[%{FILEMODES:perms} %{FILENAMES}\n]' "$rpm_file" |
+        grep -v '^l' |
+        awk '$1 ~ /^.{8}w/ { print }'
+)"
+if [ -n "$writable" ]; then
+    echo "This package contains world-writable files:" >&2
+    printf '%s\n' "$writable" >&2
+    echo "Any local account could replace them. See the mode block in build/linux/dodossh.spec." >&2
+    exit 1
+fi
+
+echo "==> requires:"
+printf '%s\n' "$requires" | sed 's/^/    /'
+
+echo "==> $(basename "$rpm_file") is a $name $rpm_version package and says so."
+
+# ◆ THE MANIFEST, WHICH IS WHAT AN INSTALLED CLIENT READS AND NOT WHAT A PERSON READS.
+#
+# The same arrangement the android job publishes, for the same reason its comment gives: a release's own
+# JSON is a large document that changes shape when the forge is upgraded, and the client fetches this on a
+# six-hourly timer. A small file this repository writes is a contract; a large one the forge writes is not.
+#
+# ◆ THE VERSION IN HERE IS THE RPM VERSION, AND THAT IS THE WHOLE POINT OF THE FILE.
+#
+# LinuxUpdateChannel compares with rpmvercmp, because rpm is what will accept or refuse the install — the
+# same argument AndroidChannelManifest makes for versionCode. Publishing the SemVer instead would mean the
+# client comparing one thing and the package manager another, and they disagree exactly where it matters:
+# 0.0.1-alpha.0.143 and 0.0.1~alpha.0.143 are the same build, and only the second one sorts below 0.0.1.
+#
+# The SemVer goes in too and is used for nothing but the banner. A person asked which nightly they are on
+# should be able to answer with the string the preferences screen prints, and that screen reads the
+# assembly.
+#
+# ◆ AND THE DIGEST, WHICH THE OTHER TWO CHANNELS HAVE NO NEED OF.
+#
+# Velopack verifies what it downloads and Android refuses a package whose signature does not match the
+# installed one. Nothing about an RPM fetched over HTTPS and handed to PackageKit checks that it is the
+# file this release published — TLS authenticates the forge and says nothing about the bytes surviving a
+# proxy, a cache or a partial write. So the manifest names the digest and the client refuses anything else.
+#
+# It is not a signature and must not be read as one: it is written by the same job that writes the package,
+# so it proves integrity and not provenance. Provenance here is that the address is a constant pointing at
+# the project's own forge — ADR 0013 rule 4 — and a signed repository is what would improve on it.
+manifest="$out/linux-$channel.json"
+digest="$(sha256sum "$rpm_file" | cut -d' ' -f1)"
+
+cat > "$manifest" <<JSON
+{
+  "rpmVersion": "$rpm_version-$rpm_release",
+  "version": "$version",
+  "rpm": "$(basename "$rpm_file")",
+  "sha256": "$digest"
+}
+JSON
+
+echo "==> $(basename "$manifest"):"
+sed 's/^/    /' "$manifest"
+
+# The lines the calling job reads. Written to stdout as well as to GITHUB_OUTPUT when there is one, so the
+# desk use above still tells you where the file went.
+if [ -n "${GITHUB_OUTPUT:-}" ]; then
+    {
+        echo "rpm=$rpm_file"
+        echo "rpmName=$(basename "$rpm_file")"
+        echo "manifest=$manifest"
+        echo "manifestName=$(basename "$manifest")"
+    } >> "$GITHUB_OUTPUT"
+fi
+echo "$rpm_file"

scripts/release-android.ps1

@@ -169,16 +169,66 @@ Bump ApplicationVersion, commit, move the tag, and run this again.
     if (Test-Path $Staging) { Remove-Item -Recurse -Force $Staging }
     New-Item -ItemType Directory -Path $Staging | Out-Null
 
-    # No RuntimeIdentifier, so every supported ABI is packaged. An arm64-only APK will not install on a
-    # 32-bit handset or an x86_64 emulator, and this is the package strangers are pointed at.
-    & dotnet build $Project --no-restore --configuration Release `
-        -t:SignAndroidPackage `
-        -p:DodoChannel=release `
-        -p:AndroidKeyStore=true `
-        -p:AndroidSigningKeyStore=$KeystorePath `
-        -p:AndroidSigningKeyAlias=$KeyAlias `
-        -p:AndroidSigningStorePass=$storePlain `
-        -p:AndroidSigningKeyPass=$keyPlain
+    # ◆ THE PASSWORDS GO THROUGH THE ENVIRONMENT, BECAUSE -p: PUTS THEM ON A COMMAND LINE.
+    #
+    # They used to be written straight into -p:AndroidSigningStorePass=..., which reads as private and is
+    # not. A process's command line is not a secret on Windows: any process running as this user can read
+    # every argument of every other with one call —
+    #
+    #     Get-CimInstance Win32_Process | Select-Object Name, CommandLine
+    #
+    # — and process-creation auditing, EDR agents and crash dumps all record it as a matter of course. The
+    # comment above says the passwords are never stored, and with the arguments written that way it was
+    # true of the disk and false of everything watching the machine.
+    #
+    # env: is .NET for Android's own answer to exactly this, and it names the variable rather than carrying
+    # the value: MSBuild is handed the string "env:DODOSSH_ANDROID_STOREPASS" and reads the password out of
+    # the environment itself. The one restriction is $(AndroidPackageFormat) = aab, where env: is not
+    # supported and file: is the documented alternative; this channel ships an APK, and the guard below
+    # fails loudly rather than silently signing with the literal string if that ever changes.
+    #
+    # Not a perfect channel — an environment block is readable by a process with rights over this one, the
+    # same as memory is — but it is off the command line, out of the audit trail and out of any log that
+    # echoes the build invocation, which is the whole of the difference being bought here.
+    #
+    # -nodeReuse:false so no MSBuild worker outlives this call holding the variables in its environment.
+    # The finally clears them from this process either way, and a persistent node would have kept its own
+    # inherited copy for the next quarter of an hour regardless of what happens here.
+    $packageFormat = (& dotnet msbuild $Project -getProperty:AndroidPackageFormat -nologo `
+        -p:DodoChannel=release | ForEach-Object { $_.Trim() } | Where-Object { $_ })
+
+    if ($packageFormat -eq 'aab') {
+        Stop-With @'
+This channel now builds an aab, and the env: prefix this script signs with is not supported for that
+format. The signing would silently take "env:DODOSSH_ANDROID_STOREPASS" for the password itself.
+
+Switch both passwords to the file: prefix, which is supported for aab, and make sure whatever writes
+that file removes it again.
+'@
+    }
+
+    try {
+        $env:DODOSSH_ANDROID_STOREPASS = $storePlain
+        $env:DODOSSH_ANDROID_KEYPASS = $keyPlain
+
+        # No RuntimeIdentifier, so every supported ABI is packaged. An arm64-only APK will not install on a
+        # 32-bit handset or an x86_64 emulator, and this is the package strangers are pointed at.
+        & dotnet build $Project --no-restore --configuration Release `
+            -nodeReuse:false `
+            -t:SignAndroidPackage `
+            -p:DodoChannel=release `
+            -p:AndroidKeyStore=true `
+            -p:AndroidSigningKeyStore=$KeystorePath `
+            -p:AndroidSigningKeyAlias=$KeyAlias `
+            -p:AndroidSigningStorePass=env:DODOSSH_ANDROID_STOREPASS `
+            -p:AndroidSigningKeyPass=env:DODOSSH_ANDROID_KEYPASS
+    }
+    finally {
+        # SetEnvironmentVariable rather than Remove-Item, which throws when the variable is not there —
+        # and this runs on the failure path, where it may well not be.
+        [Environment]::SetEnvironmentVariable('DODOSSH_ANDROID_STOREPASS', $null)
+        [Environment]::SetEnvironmentVariable('DODOSSH_ANDROID_KEYPASS', $null)
+    }
 
     if ($LASTEXITCODE -ne 0) { Stop-With 'The build failed.' }
 

scripts/release-macos.sh

@@ -0,0 +1,422 @@
+#!/usr/bin/env bash
+#
+# Builds, packages and publishes the macOS desktop client.
+#
+# The counterpart of scripts/release-windows.ps1, and deliberately the same shape: run by a person, on a
+# Mac that is not a CI runner, in two phases with the upload withheld until somebody has installed what
+# phase one built and walked the manual checks. docs/adr/0011-android-distribution.md rule 1 puts the
+# capability to ship somebody a build on a machine which is not a runner, and
+# docs/adr/0013-desktop-distribution-and-updates.md explains why the token that writes a Gitea release is
+# that capability: Velopack clients trust their feed and do not verify a package signature when they apply
+# it, so whoever can write a release can ship an update every install runs.
+#
+#   1. Without --upload: builds, signs, notarizes, packs, and stops. Nothing has left this machine
+#      except the notarization submission, which Apple sees and users do not.
+#   2. With --upload: asks for the forge token and publishes what phase one produced. It does not
+#      rebuild, so the bytes that reach users are the bytes that were installed and checked.
+#
+# ◆ WHAT IS DIFFERENT FROM THE WINDOWS SCRIPT, AND WHY.
+#
+# Signing is not optional here. On Windows an unsigned installer costs a SmartScreen dialog once per
+# user, which is why that script has no --signParams and says so. On macOS an un-notarized download is
+# refused outright by Gatekeeper — not warned about, refused — so the Developer ID certificate and the
+# notarization round trip are the price of the package being installable at all, not an improvement to
+# be bought later.
+#
+# ◆ CREDENTIALS COME FROM THE KEYCHAIN AND THE ENVIRONMENT, NOT FROM THIS FILE.
+#
+# Three values are read from the environment, and none of them is itself a secret — they name things the
+# keychain holds, and the keychain is what guards the private key and the App Store Connect credentials:
+#
+#   DODOSSH_SIGN_APP_IDENTITY      e.g. "Developer ID Application: DodoTech (TEAMID)"
+#   DODOSSH_SIGN_INSTALL_IDENTITY  e.g. "Developer ID Installer: DodoTech (TEAMID)"
+#   DODOSSH_NOTARY_PROFILE         the profile name given to `xcrun notarytool store-credentials`
+#
+# `security find-identity -v -p codesigning` lists the first two exactly as codesign wants them. The
+# third is created once per machine:
+#
+#   xcrun notarytool store-credentials DodoSSH \
+#     --apple-id you@example.com --team-id TEAMID --password <app-specific-password>
+#
+# The forge token is the one real secret, and it is prompted for rather than read from a file or an
+# ambient environment variable, and only in the phase that needs it — for the reason the Windows script
+# gives: the fewer minutes a credential that can publish an update spends in a shell's memory the better.
+# It reaches vpk through VPK_TOKEN, exported for that one call and unset afterwards, because the
+# alternative is --token on a command line that is not private. See the comment at the upload itself.
+#
+# Usage:
+#   bash scripts/release-macos.sh
+#   bash scripts/release-macos.sh --upload
+#   bash scripts/release-macos.sh --skip-tests
+
+set -euo pipefail
+
+UPLOAD=0
+SKIP_TESTS=0
+
+for arg in "$@"; do
+    case "$arg" in
+        --upload)     UPLOAD=1 ;;
+        --skip-tests) SKIP_TESTS=1 ;;
+        *)
+            echo "Unknown argument: $arg" >&2
+            echo "Usage: bash scripts/release-macos.sh [--upload] [--skip-tests]" >&2
+            exit 1
+            ;;
+    esac
+done
+
+# ---- The contract with every installed client ---------------------------------------------------------
+
+# Velopack's identity for this application, and it is effectively irreversible for the reasons the Windows
+# script states — it is what an installed client matches an update against.
+#
+# ◆ THE SAME PACK ID AS WINDOWS, AND ON THIS PLATFORM IT IS VISIBLE.
+#
+# vpk names the bundle after the pack id, so this produces DodoSSH.Desktop.app rather than DodoSSH.app,
+# and that is what somebody sees in /Applications. It is kept anyway, because the alternative is worse:
+# a pack id of DodoSSH would put Velopack's install and its uninstall on ~/Library/Application Support/
+# DodoSSH, which is exactly where ClientPaths keeps the encrypted cache, the outbox of changes not yet
+# pushed and the device key. Sharing that directory would mean an uninstall silently taking a user's
+# un-synced work with it. The same reasoning, and the same conclusion, as the Windows script.
+#
+# What a person actually reads is CFBundleDisplayName, which build/macos/Info.plist.template sets to
+# DodoSSH. So the bundle keeps the id and the Dock shows the product.
+PACK_ID='DodoSSH.Desktop'
+PACK_TITLE='DodoSSH'
+PACK_AUTHORS='DodoTech'
+
+# The project's own forge. Never a DodoSSH deployment — ADR 0011 rule 2. The same URL is a constant in
+# VelopackUpdateChannel, and the two have to agree or the client polls somewhere nothing is published.
+# The owner is part of it: Gitea left a 301 at the old organisation's path, which a GET follows and an
+# upload does not.
+REPO_URL='https://git.dodotech.cloud/DodoTech-Public/DodoSSH'
+
+# A contract with VelopackUpdateChannel.MacReleaseChannel. Velopack's macOS default is also "osx", so
+# leaving it unsaid on both sides would work — but unsaid here and stated there is how a feed goes quiet
+# with no error at all: the client checks, finds nothing, and reports itself up to date forever.
+CHANNEL='osx'
+
+# ◆ ARM64 ONLY, AND THAT IS A DECISION RATHER THAN AN OVERSIGHT.
+#
+# Velopack keys a channel to one architecture, so shipping Intel too means a second channel, a second
+# publish, a second set of deltas and a second thing to keep in step with the client's channel picker.
+# That is all affordable. What is not currently affordable is testing it: nobody here has an Intel Mac,
+# and docs/manual-checks.md exists because this project does not ship desktop builds no one has run.
+# An x64 package built blind and published beside a checked arm64 one would be the only artefact in this
+# repository that reached users unverified.
+#
+# Adding it later is this constant, a second channel name in VelopackUpdateChannel, and a picker keyed on
+# RuntimeInformation.ProcessArchitecture — which reports X64 for a build running under Rosetta, so an
+# Intel build correctly stays on the Intel feed. The work is small; the check is the part that is missing.
+RUNTIME='osx-arm64'
+
+REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+PROJECT="$REPO_ROOT/src/DodoSSH.Client.App/DodoSSH.Client.App.csproj"
+SOLUTION="$REPO_ROOT/DodoSSH.slnx"
+PUBLISH_DIR="$REPO_ROOT/publish/$RUNTIME"
+RELEASES_DIR="$REPO_ROOT/Releases"
+ICON="$REPO_ROOT/src/DodoSSH.Client.App/Assets/dodossh.icns"
+ENTITLEMENTS="$REPO_ROOT/build/macos/DodoSSH.entitlements"
+PLIST_TEMPLATE="$REPO_ROOT/build/macos/Info.plist.template"
+
+write_step() { printf '\n\033[36m==> %s\033[0m\n' "$1"; }
+stop_with()  { printf '\n\033[31m%s\033[0m\n' "$1" >&2; exit 1; }
+
+# ---- Is this machine able to do the job at all? -------------------------------------------------------
+
+if [ "$(uname -s)" != 'Darwin' ]; then
+    # codesign, notarytool and stapler are Apple tooling and exist nowhere else. The build and even the
+    # .app bundle cross-compile fine from Windows or Linux — `vpk [osx] bundle` does exactly that, and
+    # ci.yml uses it to prove the bundle still builds — but a signed, notarized, installable package
+    # cannot be produced anywhere but here.
+    stop_with 'This builds a signed macOS package and has to run on macOS.'
+fi
+
+for tool in dotnet git xcrun codesign; do
+    command -v "$tool" >/dev/null 2>&1 || stop_with "$tool is not on PATH."
+done
+
+# Checked before anything is built rather than at the step that uses them. Notarization is the last thing
+# this script does and the slowest, and discovering there that a profile name was never exported means
+# throwing away a full build and test run.
+for required in DODOSSH_SIGN_APP_IDENTITY DODOSSH_SIGN_INSTALL_IDENTITY DODOSSH_NOTARY_PROFILE; do
+    if [ -z "${!required-}" ]; then
+        stop_with "$required is not set. See the header of this script for what the three are and how to make them."
+    fi
+done
+
+cd "$REPO_ROOT"
+
+# ---- What is being released ---------------------------------------------------------------------------
+
+# Restored before the version is read, and both halves are load-bearing — the same two traps the Windows
+# script documents. -t:MinVer, because -getProperty alone evaluates the project and runs no targets, while
+# MinVer sets Version from inside one, so the read would answer the SDK's default 1.0.0 regardless of the
+# tag. And a restore first, because naming a target that arrives with a package fails MSB4057 on a clean
+# clone where obj/ has no MinVer targets to import yet.
+write_step 'Restoring the desktop head, so the version can be read'
+dotnet restore "$PROJECT" --locked-mode || stop_with 'Restore failed.'
+
+VERSION="$(dotnet msbuild "$PROJECT" -getProperty:Version -t:MinVer -nologo | tr -d '[:space:]')"
+[ -n "$VERSION" ] || stop_with 'Could not read the version from MSBuild.'
+
+TAG="v$VERSION"
+
+# Apple's two version keys take one to three dot-separated integers and nothing else, so a prerelease
+# version has to have its suffix removed before it reaches the plist. 1.2.3-rc.1 becomes 1.2.3.
+#
+# The full version, suffix and all, is what vpk packs and what the release index carries, so the updater
+# still tells an rc from the release it precedes. These two keys are for Finder and Gatekeeper, which
+# care that the string parses and not what it says. See build/macos/Info.plist.template.
+PLIST_VERSION="${VERSION%%-*}"
+PLIST_VERSION="${PLIST_VERSION%%+*}"
+
+write_step "DodoSSH $VERSION  ($PACK_ID, channel $CHANNEL, $RUNTIME)"
+
+# ---- Phase 2: publish what phase 1 built --------------------------------------------------------------
+
+if [ "$UPLOAD" -eq 1 ]; then
+    # The installer package is the artefact a person downloads, so its absence is the honest test of
+    # whether phase one ever ran. A directory holding only a .nupkg is a pack that failed part way.
+    if ! ls "$RELEASES_DIR"/*.pkg >/dev/null 2>&1; then
+        stop_with "Nothing to upload: $RELEASES_DIR has no .pkg. Run this without --upload first."
+    fi
+
+    echo "About to publish the contents of $RELEASES_DIR to $REPO_URL as $TAG."
+    echo 'Only do this once you have installed it and walked Phase 18 of docs/manual-checks.md.'
+
+    # -s so the token is never echoed and never lands in the shell's history.
+    printf 'Gitea token (write:repository): '
+    read -r -s TOKEN
+    echo
+
+    [ -n "$TOKEN" ] || stop_with 'No token given.'
+
+    # --merge because Gitea already has a release entry for the pushed tag — and on this platform it may
+    # also already hold the Windows package for the same tag, which is the case --merge is really doing
+    # the work for: without it the second platform to publish a given version fails on a release that
+    # exists, and with it the two sit side by side under one tag. --channel keeps the indexes apart.
+    #
+    # No --token in this array, and that is deliberate. It used to be there, and an argument is not a
+    # private place to put a credential: /proc/<pid>/cmdline is world-readable on Linux, `ps` hands out
+    # other processes' arguments freely, and whatever agent is auditing process creation on this machine
+    # keeps its own copy. The comment above is about shell history and was never the whole story.
+    #
+    # vpk documents VPK_TOKEN as --token's environment fallback, so the value never becomes an argument.
+    # Exported for the one call and unset straight after; an environment block is still readable by a
+    # process with rights over this one, but it is out of every process listing and audit log.
+    UPLOAD_ARGS=(
+        upload gitea
+        --repoUrl "$REPO_URL"
+        --outputDir "$RELEASES_DIR"
+        --channel "$CHANNEL"
+        --releaseName "$TAG"
+        --tag "$TAG"
+        --merge
+        --publish
+    )
+
+    # Mirrors the rule the docker image job and the Windows script already apply to the same tag, so a
+    # release candidate is a prerelease in every channel or in none.
+    case "$VERSION" in
+        *-*) UPLOAD_ARGS+=(--pre) ;;
+    esac
+
+    write_step 'Uploading'
+
+    # Unset on the way out whether or not the upload worked, so a failed release does not leave the token
+    # sitting in this shell for whatever runs next.
+    export VPK_TOKEN="$TOKEN"
+    vpk_status=0
+    dotnet vpk "${UPLOAD_ARGS[@]}" || vpk_status=$?
+    unset VPK_TOKEN
+
+    [ "$vpk_status" -eq 0 ] || stop_with 'vpk upload failed.'
+
+    write_step "Published $TAG."
+    exit 0
+fi
+
+# ---- Phase 1: build, sign, notarize, pack -------------------------------------------------------------
+
+[ -z "$(git status --porcelain)" ] || stop_with 'The working tree is not clean. A release is cut from a commit, not from a desk.'
+
+HEAD_TAG="$(git describe --exact-match --tags HEAD 2>/dev/null || true)"
+[ -n "$HEAD_TAG" ] || stop_with "HEAD is not tagged. Tag it $TAG first, or change the version and tag that."
+
+# Cannot happen while MinVer is deriving the version from this very tag, and checked anyway: the day
+# somebody pins a version by hand this is the guard that notices.
+[ "$HEAD_TAG" = "$TAG" ] || stop_with "HEAD is tagged $HEAD_TAG but the computed version is $VERSION."
+
+write_step 'Restoring tools'
+dotnet tool restore || stop_with 'dotnet tool restore failed.'
+
+write_step 'Restoring packages (locked, exactly as CI does)'
+dotnet restore "$SOLUTION" --locked-mode || stop_with 'Restore failed. A lock file that only works on Linux fails here.'
+
+write_step 'Building'
+dotnet build "$SOLUTION" --no-restore --configuration Release || stop_with 'Build failed.'
+
+if [ "$SKIP_TESTS" -eq 0 ]; then
+    # The end-to-end suite starts containers and takes minutes. It is run here anyway rather than taken
+    # on trust from CI, because a tag is the one build nobody is watching — and on this platform there is
+    # a second reason: CI has no macOS runner, so this is the only place the suite ever runs on a Mac at
+    # all. Everything docs/platform-flags.md lists as unverified on macOS is verified here or nowhere.
+    write_step 'Testing'
+    dotnet test "$SOLUTION" --no-build --configuration Release || stop_with 'Tests failed.'
+fi
+
+write_step "Publishing $RUNTIME"
+rm -rf "$PUBLISH_DIR"
+
+# Self-contained, and not single-file, for the reasons the Windows script gives: the native libraries ship
+# per RID and a self-extracting bundle breaks delta updates.
+#
+# RestoreLockedMode=false, and the lock files put back straight afterwards. A RID-specific publish resolves
+# a graph the committed lock files do not describe, because they are deliberately kept RID-free —
+# declaring a RID on the head writes a net10.0/<rid> target into every project it references transitively,
+# including DodoSSH.Contracts and DodoSSH.Crypto, and the API's Dockerfile then restores those with no RID
+# under locked mode and fails NU1004. Packaging the desktop client would have broken the server's image
+# build. The gate that matters is the locked solution restore above, which is untouched.
+dotnet publish "$PROJECT" \
+    --configuration Release \
+    --runtime "$RUNTIME" \
+    --self-contained true \
+    --output "$PUBLISH_DIR" \
+    -p:RestoreLockedMode=false \
+    || stop_with 'Publish failed.'
+
+# An unlocked restore rewrites the lock files it walked. Left there, the next commit would carry exactly
+# the change that breaks the image build. Safe to do bluntly because this script refuses to run on a dirty
+# tree, so anything modified here is its own.
+git checkout -- '*packages.lock.json' || stop_with 'Could not restore the lock files after publishing.'
+
+# Checked rather than assumed. A publish directory without Velopack.dll would pack into an installer for an
+# application that never checks for updates — which looks completely normal until the next release goes out
+# and nobody receives it.
+for required in DodoSSH Velopack.dll; do
+    [ -e "$PUBLISH_DIR/$required" ] || stop_with "$required is missing from $PUBLISH_DIR."
+done
+
+echo "  $(du -sh "$PUBLISH_DIR" | cut -f1) in $(find "$PUBLISH_DIR" -type f | wc -l | tr -d ' ') files"
+
+# ---- Signing the native libraries, before vpk signs anything ------------------------------------------
+
+# ◆ THIS LOOP IS WHY NOTARIZATION SUCCEEDS, AND IT LOOKS REDUNDANT.
+#
+# vpk signs the finished bundle itself, with `codesign -f -v --timestamp --options runtime --entitlements
+# <file> --deep`, and --deep is documented by Apple as the wrong way to sign nested code. Apple's guidance
+# is inside-out: sign each nested binary first, then the bundle around it. --deep does the reverse in one
+# pass and applies the outer entitlements to everything it touches.
+#
+# In practice --deep alone is where the failure recorded in docs/platform-flags.md comes from — a
+# notarization rejection that does not name the offending file, on a submission that took its time getting
+# there. Signing each dylib properly first means vpk's pass has nothing left to get wrong, and re-signing
+# an already correctly signed binary with -f is a no-op in effect.
+#
+# No --entitlements here, and that is the difference that matters. Entitlements belong on the main
+# executable; a dylib carrying allow-jit is at best meaningless and at worst a rejection.
+write_step 'Signing native libraries'
+
+# createdump is a Mach-O executable the runtime ships and it is signed like the libraries: a nested
+# executable that is not signed fails notarization exactly as an unsigned dylib does, and it is the one
+# people forget because it has no extension to grep for.
+NATIVE_COUNT=0
+while IFS= read -r -d '' binary; do
+    codesign --force --verbose=0 --timestamp --options runtime \
+        --sign "$DODOSSH_SIGN_APP_IDENTITY" "$binary" \
+        || stop_with "codesign failed on $binary"
+    NATIVE_COUNT=$((NATIVE_COUNT + 1))
+done < <(find "$PUBLISH_DIR" \( -name '*.dylib' -o -name 'createdump' \) -type f -print0)
+
+[ "$NATIVE_COUNT" -gt 0 ] || stop_with "No native binaries found under $PUBLISH_DIR, which cannot be right for a self-contained publish."
+echo "  signed $NATIVE_COUNT native binaries"
+
+# ---- The bundle's Info.plist --------------------------------------------------------------------------
+
+# Rendered rather than committed, because vpk copies a custom plist verbatim and substitutes nothing —
+# so a committed one would carry whatever version it was written with into every release afterwards.
+# See the header of build/macos/Info.plist.template.
+write_step "Rendering Info.plist for $PLIST_VERSION"
+RENDERED_PLIST="$(mktemp -t dodossh-plist)"
+trap 'rm -f "$RENDERED_PLIST"' EXIT
+
+sed "s/@VERSION@/$PLIST_VERSION/g" "$PLIST_TEMPLATE" > "$RENDERED_PLIST"
+
+# The placeholder is the whole mechanism, so its absence is checked rather than hoped for. A template
+# somebody edited into a literal version would otherwise sail through and pin every future release to it.
+grep -q '@VERSION@' "$PLIST_TEMPLATE" || stop_with "$PLIST_TEMPLATE has no @VERSION@ placeholder left in it."
+! grep -q '@VERSION@' "$RENDERED_PLIST" || stop_with 'Substitution into the rendered Info.plist did not take.'
+
+mkdir -p "$RELEASES_DIR"
+
+# The previous release, so a delta can be built against it. Tolerated when it finds nothing: the first
+# macOS release has no predecessor, and a hard failure here would make cutting it impossible.
+write_step 'Fetching the previous release, for deltas'
+if ! dotnet vpk download gitea --repoUrl "$REPO_URL" --outputDir "$RELEASES_DIR" --channel "$CHANNEL"; then
+    echo '  Nothing came down. This package will be full-only, which is right for a first release.'
+fi
+
+# ---- Pack, sign, notarize, staple ---------------------------------------------------------------------
+
+# One command does the rest, and it is worth knowing what it is doing on your behalf, because the slow
+# part is not local: it builds the .app from the published files, signs it with the Developer ID
+# certificate and the entitlements below, submits it to Apple with `xcrun notarytool submit --wait`,
+# staples the resulting ticket to the package, and then builds the .pkg installer and the release index.
+#
+# The notarization wait is the reason this step can take a quarter of an hour and occasionally much
+# longer — it is a queue at Apple, not a computation here, and vpk's own message says so.
+#
+# --signInstallIdentity is a different certificate from --signAppIdentity, and the pair is not
+# interchangeable: "Developer ID Application" signs the bundle, "Developer ID Installer" signs the .pkg.
+# Passing one where the other belongs fails with a message about an identity that cannot be found, which
+# reads like a keychain problem rather than like the wrong certificate.
+write_step 'Packing, signing and notarizing (the notarization wait is Apple queueing, not this machine)'
+
+dotnet vpk pack \
+    --packId "$PACK_ID" \
+    --packVersion "$VERSION" \
+    --packDir "$PUBLISH_DIR" \
+    --packTitle "$PACK_TITLE" \
+    --packAuthors "$PACK_AUTHORS" \
+    --mainExe 'DodoSSH' \
+    --icon "$ICON" \
+    --plist "$RENDERED_PLIST" \
+    --entitlements "$ENTITLEMENTS" \
+    --signAppIdentity "$DODOSSH_SIGN_APP_IDENTITY" \
+    --signInstallIdentity "$DODOSSH_SIGN_INSTALL_IDENTITY" \
+    --notaryProfile "$DODOSSH_NOTARY_PROFILE" \
+    --runtime "$RUNTIME" \
+    --channel "$CHANNEL" \
+    --outputDir "$RELEASES_DIR" \
+    || stop_with 'vpk pack failed.'
+
+# ---- Did the notarization actually take? --------------------------------------------------------------
+
+# Asked rather than assumed, and this is the check worth having above all the others. A package whose
+# ticket did not staple is indistinguishable from a good one on the machine that built it — the Mac that
+# signed something trusts it locally — and reveals itself only on somebody else's machine, as a refusal
+# to open at all. spctl assesses it the way Gatekeeper will on a machine that has never seen this
+# certificate.
+write_step 'Verifying the notarization the way another Mac will'
+
+PKG="$(ls -t "$RELEASES_DIR"/*.pkg 2>/dev/null | head -n 1)"
+[ -n "$PKG" ] || stop_with 'vpk pack reported success but produced no .pkg.'
+
+if ! spctl --assess --type install --verbose=4 "$PKG"; then
+    stop_with "Gatekeeper rejects $PKG. It is signed but the notarization ticket is missing or stale; do not upload it."
+fi
+
+xcrun stapler validate "$PKG" || stop_with "The notarization ticket is not stapled to $PKG."
+
+write_step 'Built, notarized, and deliberately not uploaded'
+
+ls -lh "$RELEASES_DIR" | tail -n +2
+
+cat <<EOF
+
+Next:
+  1. Install the .pkg above and walk Phase 18 of docs/manual-checks.md.
+  2. Then: bash scripts/release-macos.sh --upload
+EOF

scripts/release-windows.ps1

@@ -16,9 +16,11 @@
       2. With -Upload: asks for the forge token and publishes what phase 1 produced. It does not rebuild,
          so the bytes that reach users are the bytes that were installed and checked.
 
-    The token is prompted for rather than read from a file or an environment variable, and only in the phase
-    that needs it — the build does not, and the fewer minutes a credential that can publish an update spends
-    in a shell's memory the better.
+    The token is prompted for rather than read from a file or an ambient environment variable, and only in
+    the phase that needs it — the build does not, and the fewer minutes a credential that can publish an
+    update spends in a shell's memory the better. It reaches vpk through VPK_TOKEN, set for that one call
+    and cleared afterwards, because the alternative is --token on a command line every process on this
+    machine can read. See the comment at the upload itself.
 
 .PARAMETER Upload
     Publish the packages already in Releases/ instead of building.
@@ -71,6 +73,11 @@ $RepoUrl = 'https://git.dodotech.cloud/DodoTech-Public/DodoSSH'
 # A contract with VelopackUpdateChannel.ReleaseChannel. It is Velopack's Windows default, so leaving it
 # unsaid on both sides would work too — but unsaid here and stated there is how a feed goes quiet with no
 # error at all: the client checks, finds nothing, and reports itself up to date forever.
+#
+# The nightly channel is the same three values with different contents — DodoSSH.Desktop.Nightly, DodoSSH
+# Nightly, win-nightly — and they live in the `desktop nightly` job in .github/workflows/ci.yml rather than
+# here, because that build is CI's and this script is a person's. Nothing shares them, deliberately: this
+# file must keep working if that job is deleted. See ADR 0013 decision 9 for what the separation buys.
 $Channel = 'win'
 
 $RepoRoot = Split-Path -Parent $PSScriptRoot
@@ -100,11 +107,20 @@ Push-Location $RepoRoot
 try {
     # ---- What is being released -----------------------------------------------------------------------
 
-    # -t:MinVer, and it is load-bearing. -getProperty on its own evaluates the project and runs no
-    # targets, while MinVer sets Version from inside one — so this read answered the SDK's default
-    # 1.0.0 regardless of the tag, and the check below would then have refused to build anything not
-    # tagged v1.0.0. CI's tag check had the same line and the same fault; both are fixed, and both
-    # say so, because this is the version that ends up in the package a client compares against.
+    # Restored before the version is read, and both halves of that sentence are load-bearing.
+    #
+    # -t:MinVer, because -getProperty on its own evaluates the project and runs no targets, while MinVer
+    # sets Version from inside one — so this read answered the SDK's default 1.0.0 regardless of the tag,
+    # and the check below would then have refused to build anything not tagged v1.0.0. CI's tag check had
+    # the same line and the same fault.
+    #
+    # And a restore first, because naming a target that arrives with a package fails MSB4057 on a clean
+    # clone, where obj/ has no MinVer targets to import yet. It costs seconds on a machine that has built
+    # before, which is every machine except the one this would otherwise fail on.
+    Write-Step 'Restoring the desktop head, so the version can be read'
+    & dotnet restore $Project --locked-mode
+    if ($LASTEXITCODE -ne 0) { Stop-With 'Restore failed.' }
+
     $version = (& dotnet msbuild $Project -getProperty:Version -t:MinVer -nologo) -replace '\s', ''
     if ([string]::IsNullOrWhiteSpace($version)) {
         Stop-With 'Could not read the version from MSBuild.'
@@ -138,10 +154,24 @@ try {
         # --merge because Gitea already has a release entry for the pushed tag, and without it the upload
         # fails on a release that exists. --pre mirrors the rule the docker image job already applies to the
         # same tag, so a release candidate is a prerelease in both channels or in neither.
+        #
+        # ◆ NO --token HERE. IT WAS AN ARGUMENT, AND AN ARGUMENT IS PUBLIC ON THIS MACHINE.
+        #
+        # The comment above says the token never lands in shell history, which was true and was not the
+        # whole story: it went straight into vpk's command line, and any process running as this user can
+        # read that off a running process with one call —
+        #
+        #     Get-CimInstance Win32_Process | Select-Object Name, CommandLine
+        #
+        # — with process-creation auditing and EDR agents keeping their own copy afterwards. A token that
+        # can publish an update every installed copy will fetch and run deserves better than that.
+        #
+        # vpk documents VPK_TOKEN as --token's environment fallback, so the flag simply goes away and the
+        # value never becomes an argument at all. An environment block is still readable by a process with
+        # rights over this one; what it is not is a string sitting in a process listing and an audit log.
         $uploadArgs = @(
             'upload', 'gitea',
             '--repoUrl', $RepoUrl,
-            '--token', $token,
             '--outputDir', $ReleasesDir,
             '--channel', $Channel,
             '--releaseName', $tag,
@@ -155,7 +185,17 @@ try {
         }
 
         Write-Step 'Uploading'
-        & dotnet vpk @uploadArgs
+
+        try {
+            $env:VPK_TOKEN = $token
+            & dotnet vpk @uploadArgs
+        }
+        finally {
+            # SetEnvironmentVariable rather than Remove-Item, which throws when the variable is not there —
+            # and this runs on the failure path, where it may well not be.
+            [Environment]::SetEnvironmentVariable('VPK_TOKEN', $null)
+        }
+
         if ($LASTEXITCODE -ne 0) { Stop-With 'vpk upload failed.' }
 
         Write-Step "Published $tag."

src/DodoSSH.Api/Features/Meta/MetaEndpoints.cs

@@ -26,9 +26,6 @@ internal sealed class GetMetaEndpoint(
     /// <summary>Sync semantics version. Bumped when push or pull behaviour changes.</summary>
     internal const int SyncProtocolVersion = 1;
 
-    /// <summary>Feature flag for the relay.</summary>
-    internal const string RelayFeature = "relay";
-
     /// <inheritdoc />
     public override void Configure()
     {
@@ -50,7 +47,7 @@ internal sealed class GetMetaEndpoint(
         List<string> features = ["teams"];
         if (relay.Value.Enabled)
         {
-            features.Add(RelayFeature);
+            features.Add(RelayProtocol.Feature);
         }
 
         // Advertised so a client knows whether to hold a socket open or rely on its timer. Absence is

src/DodoSSH.Api/Features/Relay/RelayAddressPolicy.cs

@@ -0,0 +1,192 @@
+using System.Net;
+using System.Net.Sockets;
+using DodoSSH.Api.Setup;
+
+namespace DodoSSH.Api.Features.Relay;
+
+/// <summary>
+/// Decides which addresses and ports the relay may dial. ADR 0004 §3, in the order it is written.
+/// </summary>
+/// <remarks>
+/// <para>
+/// This runs on <em>addresses</em>, never on names. The name is resolved once, at ticket issue, and what
+/// this permits goes into the ticket; the socket dials those addresses and nothing else. A check on the
+/// name would be defeated by DNS rebinding — answer with a public address for the check and a loopback
+/// one for the dial — which is the attack ADR 0004 is built around.
+/// </para>
+/// <para>
+/// Three layers, and the order is the security property. The <see cref="Refused"/> list comes first and
+/// has no configuration at all: nothing an operator sets, and nothing a host row says, reaches a
+/// loopback, link-local, unspecified, multicast or metadata address. Then the operator's own denials,
+/// then their allow-list, then the private-network switch.
+/// </para>
+/// <para>
+/// <b>An address that embeds another is judged as the one it embeds.</b> <c>::ffff:127.0.0.1</c> is
+/// loopback however it is spelled, and a check that compared it against IPv6 loopback alone would pass
+/// it — the classic bypass the ADR names. The NAT64 well-known prefix is unwrapped the same way, because
+/// on a network with a NAT64 gateway <c>64:ff9b::a9fe:a9fe</c> <em>is</em> the metadata endpoint.
+/// </para>
+/// </remarks>
+internal sealed class RelayAddressPolicy
+{
+    /// <summary>
+    /// Addresses the relay never dials, whatever it is configured to do.
+    /// </summary>
+    /// <remarks>
+    /// Broader than the ADR's list in two places, both deliberately. <c>240.0.0.0/4</c> — reserved, and
+    /// containing the limited broadcast address — has no host worth reaching. And IPv6 <c>::/96</c>, the
+    /// deprecated IPv4-compatible block, covers the unspecified address and loopback while also refusing
+    /// the spellings of IPv4 that a stack may or may not map; nothing legitimate lives there.
+    /// </remarks>
+    private static readonly IPNetwork[] Refused =
+    [
+        // IPv4: "this network", loopback, link-local (which holds 169.254.169.254), multicast, reserved.
+        IPNetwork.Parse("0.0.0.0/8"),
+        IPNetwork.Parse("127.0.0.0/8"),
+        IPNetwork.Parse("169.254.0.0/16"),
+        IPNetwork.Parse("224.0.0.0/4"),
+        IPNetwork.Parse("240.0.0.0/4"),
+
+        // Alibaba Cloud's metadata endpoint, which sits inside carrier-grade NAT space and so would
+        // otherwise be permitted as private — the range a Tailnet also uses, which is why the whole
+        // block cannot simply be refused.
+        IPNetwork.Parse("100.100.100.200/32"),
+
+        // IPv6: IPv4-compatible (with unspecified and loopback), link-local, multicast, and the AWS
+        // metadata endpoint's IPv6 address.
+        IPNetwork.Parse("::/96"),
+        IPNetwork.Parse("fe80::/10"),
+        IPNetwork.Parse("ff00::/8"),
+        IPNetwork.Parse("fd00:ec2::254/128"),
+    ];
+
+    /// <summary>
+    /// Ranges <see cref="RelayOptions.AllowPrivateNetworks"/> governs.
+    /// </summary>
+    /// <remarks>
+    /// Carrier-grade NAT is here rather than treated as public, because it is where a Tailnet lives and
+    /// "the server is on the same Tailnet" is one of the shapes the relay exists for. Site-local
+    /// <c>fec0::/10</c> is deprecated and was private while it lived.
+    /// </remarks>
+    private static readonly IPNetwork[] Private =
+    [
+        IPNetwork.Parse("10.0.0.0/8"),
+        IPNetwork.Parse("172.16.0.0/12"),
+        IPNetwork.Parse("192.168.0.0/16"),
+        IPNetwork.Parse("100.64.0.0/10"),
+        IPNetwork.Parse("fc00::/7"),
+        IPNetwork.Parse("fec0::/10"),
+    ];
+
+    /// <summary>The NAT64 well-known prefix, RFC 6052, whose low 32 bits are an IPv4 address.</summary>
+    private static readonly IPNetwork Nat64 = IPNetwork.Parse("64:ff9b::/96");
+
+    /// <summary>
+    /// Ports the relay never dials.
+    /// </summary>
+    /// <remarks>
+    /// SMTP and nothing else, per ADR 0004: a relay that reached port 25 is a spam cannon with the
+    /// operator's address on it. Database ports stay open, because forwarding to one is a headline use,
+    /// and the control that matters there is who may dial the host at all.
+    /// </remarks>
+    private static readonly int[] RefusedPorts = [25, 465, 587];
+
+    private readonly bool allowPrivate;
+    private readonly IPNetwork[] allowed;
+    private readonly IPNetwork[] denied;
+
+    public RelayAddressPolicy(RelayOptions options)
+    {
+        ArgumentNullException.ThrowIfNull(options);
+
+        allowPrivate = options.AllowPrivateNetworks;
+        allowed = ParseAll(options.AllowedNetworks);
+        denied = ParseAll(options.DeniedNetworks);
+    }
+
+    /// <summary>Whether a port may be dialled at all.</summary>
+    internal static bool PermitsPort(int port) =>
+        port is >= 1 and <= 65535 && !RefusedPorts.Contains(port);
+
+    /// <summary>
+    /// Whether every entry is a network this policy can read.
+    /// </summary>
+    /// <remarks>
+    /// For startup validation. A typo in a deny list that was silently skipped would be a hole the
+    /// operator believes is closed, so an unreadable entry stops the server instead.
+    /// <para>
+    /// Host bits are refused as well, although <see cref="IPNetwork.TryParse(string?, out IPNetwork)"/>
+    /// would mask them off quietly. <c>10.20.5.7/16</c> could be a /16 written from memory or a /32 with
+    /// the wrong suffix, and in a deny list those differ by sixty-five thousand addresses.
+    /// </para>
+    /// </remarks>
+    internal static bool AreNetworks(IEnumerable<string> entries) =>
+        entries.All(entry =>
+            IPNetwork.TryParse(entry, out var network)
+            && IPAddress.TryParse(entry.AsSpan(0, entry.IndexOf('/', StringComparison.Ordinal)), out var written)
+            && written.Equals(network.BaseAddress));
+
+    /// <summary>
+    /// The address the relay would actually be talking to, with any embedding unwrapped.
+    /// </summary>
+    internal static IPAddress Normalise(IPAddress address)
+    {
+        ArgumentNullException.ThrowIfNull(address);
+
+        if (address.IsIPv4MappedToIPv6)
+        {
+            return address.MapToIPv4();
+        }
+
+        if (address.AddressFamily == AddressFamily.InterNetworkV6 && Nat64.Contains(address))
+        {
+            var bytes = address.GetAddressBytes();
+            return new IPAddress(bytes.AsSpan(12, 4));
+        }
+
+        // A scope id is how an address names an interface on this machine, which is not something a
+        // host row gets to choose. Dropped, so what is judged is what is dialled.
+        return address.AddressFamily == AddressFamily.InterNetworkV6 && address.ScopeId != 0
+            ? new IPAddress(address.GetAddressBytes())
+            : address;
+    }
+
+    /// <summary>Whether the relay may dial this address.</summary>
+    /// <remarks>
+    /// Judges the normalised address. Callers should dial <see cref="Normalise"/>'s answer too, so that
+    /// the address checked and the address connected to are the same object rather than two spellings
+    /// the operating system is trusted to agree about.
+    /// </remarks>
+    internal bool Permits(IPAddress address)
+    {
+        var target = Normalise(address);
+
+        if (Refused.Any(network => Within(network, target)))
+        {
+            return false;
+        }
+
+        if (denied.Any(network => Within(network, target)))
+        {
+            return false;
+        }
+
+        if (allowed.Length > 0)
+        {
+            return allowed.Any(network => Within(network, target));
+        }
+
+        return allowPrivate || !Private.Any(network => Within(network, target));
+    }
+
+    /// <remarks>
+    /// <see cref="IPNetwork.Contains"/> answers false across address families rather than throwing, so
+    /// a list mixing IPv4 and IPv6 needs no sorting — but it is spelled out here because the one-line
+    /// alternative reads as though it compared across families.
+    /// </remarks>
+    private static bool Within(IPNetwork network, IPAddress address) =>
+        network.BaseAddress.AddressFamily == address.AddressFamily && network.Contains(address);
+
+    private static IPNetwork[] ParseAll(IEnumerable<string> entries) =>
+        [.. entries.Select(entry => IPNetwork.Parse(entry))];
+}

src/DodoSSH.Api/Features/Relay/RelayConnectEndpoint.cs

@@ -0,0 +1,722 @@
+using System.Buffers;
+using System.Net;
+using System.Net.WebSockets;
+using DodoSSH.Api.Setup;
+using DodoSSH.Contracts;
+using DodoSSH.Domain;
+using DodoSSH.Infrastructure;
+using FastEndpoints;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.Options;
+using Npgsql;
+
+namespace DodoSSH.Api.Features.Relay;
+
+/// <summary>
+/// The relay socket: a ticket in, raw bytes to one host and back.
+/// </summary>
+/// <remarks>
+/// <para>
+/// Anonymous at the HTTP layer, because the ticket is the authorization and it is the whole of it. A
+/// ticket carries no API authority — it names one host, one set of addresses and one session — so this
+/// endpoint can accept it from a client that cannot set an <c>Authorization</c> header on an upgrade,
+/// which is the constraint ADR 0004 was written around. It is verified, checked for freshness and
+/// redeemed before the upgrade is accepted, so a refused ticket costs a problem document and no socket.
+/// </para>
+/// <para>
+/// <b>Nothing here is scoped to the request.</b> A session lasts as long as a shell does, which is hours,
+/// and anything scoped held across it is a pooled database connection held across it. The database is
+/// touched three times — redeem, open the audit row, close it — each from a scope of its own that is
+/// disposed before the next byte moves. That is ADR 0004's consequence about <c>DbContext</c> lifetime.
+/// </para>
+/// <para>
+/// Backpressure is the await. Each direction reads one buffer, writes it, and does not read again until
+/// the write has completed, so a slow reader on either side stops this side reading, which fills the
+/// kernel's receive window, which throttles the sender at the far end. Nothing is queued in between, and
+/// a session costs its two buffers whatever the peers do.
+/// </para>
+/// </remarks>
+internal sealed class RelayConnectEndpoint(
+    IServiceScopeFactory scopes,
+    IRelayNetwork network,
+    RelayTicketCodec codec,
+    RelaySessionRegistry sessions,
+    IOptions<RelayOptions> options,
+    TimeProvider clock,
+    IHostApplicationLifetime lifetime,
+    ILogger<RelayConnectEndpoint> logger)
+    : EndpointWithoutRequest
+{
+    /// <summary>
+    /// One direction's buffer.
+    /// </summary>
+    /// <remarks>
+    /// Room for an SSH packet at the size OpenSSH sends them, so a bulk transfer moves in one read and
+    /// one write per packet rather than several.
+    /// </remarks>
+    private const int BufferBytes = 32 * 1024;
+
+    /// <summary>How long a close handshake, or a peer's last write, may take before the socket is dropped.</summary>
+    private static readonly TimeSpan CloseTimeout = TimeSpan.FromSeconds(5);
+
+    /// <inheritdoc />
+    public override void Configure()
+    {
+        Get(RelayProtocol.ConnectPath);
+
+        // The ticket is the authorization. See the remarks on this class, and ADR 0004.
+        AllowAnonymous();
+
+        Description(b => b
+            .WithName("RelayConnect")
+            .WithSummary("Pipes raw bytes between a WebSocket and the host a relay ticket names.")
+            .WithTags("Relay"));
+    }
+
+    /// <inheritdoc />
+    public override async Task HandleAsync(CancellationToken ct)
+    {
+        var ticket = await AdmitTicketAsync(ct).ConfigureAwait(false);
+
+        if (ticket is null)
+        {
+            return;
+        }
+
+        var lease = sessions.TryAdmit(ticket.UserId);
+
+        try
+        {
+            using var socket = await HttpContext.WebSockets
+                .AcceptWebSocketAsync(new WebSocketAcceptContext { SubProtocol = RelayProtocol.SubProtocol })
+                .ConfigureAwait(false);
+
+            if (lease is null)
+            {
+                RelayLog.LimitReached(logger, ticket.UserId, sessions.Count);
+
+                await CloseAsync(
+                        socket,
+                        new Ending(
+                            RelayCloseReason.Unspecified,
+                            RelayProtocol.TooManySessionsCloseCode,
+                            "Too many relay sessions are open."))
+                    .ConfigureAwait(false);
+
+                return;
+            }
+
+            if (lifetime.ApplicationStopping.IsCancellationRequested)
+            {
+                await CloseAsync(socket, ShuttingDown).ConfigureAwait(false);
+
+                return;
+            }
+
+            await RelayAsync(socket, ticket, ct).ConfigureAwait(false);
+        }
+        finally
+        {
+            // Inside a try that starts before the upgrade, for the reason VaultEventsEndpoint gives: an
+            // accept that throws would otherwise hold a place against this user's cap forever.
+            lease?.Dispose();
+        }
+    }
+
+    private static Ending ShuttingDown { get; } = new(
+        RelayCloseReason.ServerShutdown,
+        (int)WebSocketCloseStatus.EndpointUnavailable,
+        "The server is shutting down.");
+
+    /// <summary>
+    /// Answers everything that is not a redeemable ticket, as ordinary HTTP, before any socket exists.
+    /// </summary>
+    /// <returns>The redeemed ticket, or null when a response has been sent.</returns>
+    /// <remarks>
+    /// A ticket that is forged, expired, spent, or simply absent all answer 401
+    /// <c>relay-ticket-invalid</c>, with one remedy — ask for another — and no hint of which it was.
+    /// </remarks>
+    private async Task<RelayTicket?> AdmitTicketAsync(CancellationToken cancellationToken)
+    {
+        if (!options.Value.Enabled)
+        {
+            await Send.ResultAsync(Problems.Coded(
+                    StatusCodes.Status404NotFound,
+                    ProblemCodes.RelayUnavailable,
+                    "This server does not relay. GET /api/v1/meta lists the features it does offer."))
+                .ConfigureAwait(false);
+
+            return null;
+        }
+
+        var offered = HttpContext.WebSockets.WebSocketRequestedProtocols;
+
+        if (!HttpContext.WebSockets.IsWebSocketRequest
+            || !offered.Contains(RelayProtocol.SubProtocol, StringComparer.Ordinal))
+        {
+            await Send.ResultAsync(Problems.Coded(
+                    StatusCodes.Status400BadRequest,
+                    ProblemCodes.MalformedRequest,
+                    "This endpoint is a WebSocket. Send an upgrade request offering the "
+                    + $"'{RelayProtocol.SubProtocol}' subprotocol and the ticket as a second one, "
+                    + $"'{RelayProtocol.TicketProtocolPrefix}<ticket>'."))
+                .ConfigureAwait(false);
+
+            return null;
+        }
+
+        var sealedTicket = offered
+            .FirstOrDefault(protocol => protocol.StartsWith(RelayProtocol.TicketProtocolPrefix, StringComparison.Ordinal))
+            ?[RelayProtocol.TicketProtocolPrefix.Length..];
+
+        var ticket = sealedTicket is { Length: > 0 } ? codec.TryOpen(sealedTicket) : null;
+
+        if (ticket is null || !await RedeemAsync(ticket, cancellationToken).ConfigureAwait(false))
+        {
+            RelayLog.TicketRefused(logger, ticket?.SessionId);
+
+            await Send.ResultAsync(Problems.Coded(
+                    StatusCodes.Status401Unauthorized,
+                    ProblemCodes.RelayTicketInvalid,
+                    "This relay ticket is not valid: it is expired, already used, or was not issued here. "
+                    + $"Ask for another from POST {RelayProtocol.TicketsPath}."))
+                .ConfigureAwait(false);
+
+            return null;
+        }
+
+        return ticket;
+    }
+
+    /// <summary>
+    /// Spends the ticket, or answers false when it has been spent already.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// A unique insert, which is the only form of "once" that holds across nodes: two redemptions racing
+    /// on two nodes both attempt the insert and the database lets exactly one of them win.
+    /// </para>
+    /// <para>
+    /// Rows for tickets that can no longer verify are swept on the way. They guard nothing, and sweeping
+    /// here — on a path that runs once per session rather than once per request — keeps the table the
+    /// size of the last thirty seconds without a background job to forget about.
+    /// </para>
+    /// </remarks>
+    private async Task<bool> RedeemAsync(RelayTicket ticket, CancellationToken cancellationToken)
+    {
+        var scope = scopes.CreateAsyncScope();
+        await using var _ = scope.ConfigureAwait(false);
+
+        var database = scope.ServiceProvider.GetRequiredService<DodoDbContext>();
+        var now = clock.GetUtcNow();
+
+        await database.RelayTicketUses
+            .Where(use => use.ExpiresAtUtc < now)
+            .ExecuteDeleteAsync(cancellationToken)
+            .ConfigureAwait(false);
+
+        database.RelayTicketUses.Add(new RelayTicketUse
+        {
+            TicketId = ticket.TicketId,
+            ExpiresAtUtc = ticket.ExpiresAt,
+            UsedAtUtc = now,
+        });
+
+        try
+        {
+            await database.SaveChangesAsync(cancellationToken).ConfigureAwait(false);
+
+            return true;
+        }
+        catch (DbUpdateException exception)
+            when (string.Equals(
+                (exception.InnerException as PostgresException)?.SqlState,
+                PostgresErrorCodes.UniqueViolation,
+                StringComparison.Ordinal))
+        {
+            return false;
+        }
+    }
+
+    /// <summary>Dials the target, pipes until something ends it, and writes down what happened.</summary>
+    private async Task RelayAsync(WebSocket socket, RelayTicket ticket, CancellationToken requestAborted)
+    {
+        var started = clock.GetUtcNow();
+
+        await OpenAuditAsync(ticket, started).ConfigureAwait(false);
+
+        var dialled = await DialAsync(ticket, requestAborted).ConfigureAwait(false);
+
+        if (dialled is not var (target, address))
+        {
+            RelayLog.TargetUnreachable(logger, ticket.SessionId, ticket.HostId);
+
+            var unreachable = new Ending(
+                RelayCloseReason.TargetUnreachable,
+                RelayProtocol.TargetUnreachableCloseCode,
+                "The relay could not reach the host.");
+
+            await CloseAsync(socket, unreachable).ConfigureAwait(false);
+            await CloseAuditAsync(ticket.SessionId, null, new Counters(clock, started), unreachable.Reason).ConfigureAwait(false);
+
+            return;
+        }
+
+        var counters = new Counters(clock, started);
+        Ending ending;
+
+        await using (target.ConfigureAwait(false))
+        {
+            RelayLog.SessionOpened(logger, ticket.SessionId, ticket.UserId, ticket.HostId, sessions.Count);
+
+            ending = await PumpAsync(socket, target, counters, started, requestAborted).ConfigureAwait(false);
+        }
+
+        RelayLog.SessionClosed(
+            logger,
+            ticket.SessionId,
+            ending.Reason,
+            counters.ClientToTarget,
+            counters.TargetToClient);
+
+        await CloseAuditAsync(ticket.SessionId, address, counters, ending.Reason).ConfigureAwait(false);
+    }
+
+    /// <summary>
+    /// Connects to the first of the ticket's addresses that answers, within the connect timeout overall.
+    /// </summary>
+    /// <remarks>
+    /// Overall rather than per address, so a name with sixteen dead addresses costs five seconds rather
+    /// than eighty. In the order resolved, which is the order the operating system's resolver prefers.
+    /// </remarks>
+    private async Task<(Stream Target, IPAddress Address)?> DialAsync(
+        RelayTicket ticket,
+        CancellationToken cancellationToken)
+    {
+        using var timeout = CancellationTokenSource.CreateLinkedTokenSource(cancellationToken);
+        timeout.CancelAfter(options.Value.ConnectTimeout);
+
+        foreach (var address in ticket.TargetAddresses)
+        {
+            try
+            {
+                var target = await network
+                    .ConnectAsync(address, ticket.TargetPort, timeout.Token)
+                    .ConfigureAwait(false);
+
+                return (target, address);
+            }
+            catch (OperationCanceledException)
+            {
+                return null;
+            }
+            catch (Exception exception) when (exception is IOException or System.Net.Sockets.SocketException)
+            {
+                // The next address, if there is one.
+            }
+        }
+
+        return null;
+    }
+
+    /// <summary>
+    /// Runs both directions and the clock until one of them ends the session, then ends it politely.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// <b>No operation on the WebSocket is ever cancelled.</b> Cancelling a pending receive or send on a
+    /// .NET WebSocket aborts it, and an aborted socket cannot send the close frame that is the only way
+    /// a client learns <em>why</em> its session ended — which, for an idle timeout or a drain, is the
+    /// difference between "reconnect" and "something is broken". So the token stops the target side and
+    /// the clock; the sender is let finish its last write; the close goes out; and the receiver ends on
+    /// the client's answering close. Each wait is bounded, and the socket is aborted only when a peer
+    /// has stopped answering altogether.
+    /// </para>
+    /// </remarks>
+    private async Task<Ending> PumpAsync(
+        WebSocket socket,
+        Stream target,
+        Counters counters,
+        DateTimeOffset started,
+        CancellationToken requestAborted)
+    {
+        using var pump = CancellationTokenSource.CreateLinkedTokenSource(requestAborted);
+
+        var upstream = ClientToTargetAsync(socket, target, counters, pump.Token);
+        var downstream = TargetToClientAsync(target, socket, counters, pump.Token);
+        var minding = MindAsync(counters, started, pump.Token);
+
+        var first = await Task.WhenAny(upstream, downstream, minding).ConfigureAwait(false);
+        var ending = await first.ConfigureAwait(false) ?? new Ending(
+            RelayCloseReason.ClientClosed, (int)WebSocketCloseStatus.NormalClosure, string.Empty);
+
+        await pump.CancelAsync().ConfigureAwait(false);
+
+        // The only sender must be done before the close frame is written: a WebSocket permits one send
+        // at a time and faults permanently on a second.
+        if (!await SettlesAsync(downstream).ConfigureAwait(false))
+        {
+            socket.Abort();
+        }
+
+        await CloseAsync(socket, ending).ConfigureAwait(false);
+
+        if (!await SettlesAsync(upstream).ConfigureAwait(false))
+        {
+            socket.Abort();
+            await SettlesAsync(upstream).ConfigureAwait(false);
+        }
+
+        await SettlesAsync(minding).ConfigureAwait(false);
+
+        return ending;
+    }
+
+    /// <summary>Client to target: every binary message is the next bytes of the stream.</summary>
+    private static async Task<Ending?> ClientToTargetAsync(
+        WebSocket socket,
+        Stream target,
+        Counters counters,
+        CancellationToken cancellationToken)
+    {
+        var buffer = ArrayPool<byte>.Shared.Rent(BufferBytes);
+
+        try
+        {
+            while (true)
+            {
+                ValueWebSocketReceiveResult received;
+
+                try
+                {
+                    // Not cancellable, deliberately. See PumpAsync.
+                    received = await socket.ReceiveAsync(buffer.AsMemory(), CancellationToken.None)
+                        .ConfigureAwait(false);
+                }
+                catch (Exception exception) when (exception is WebSocketException or OperationCanceledException)
+                {
+                    return ClientGone;
+                }
+
+                if (received.MessageType == WebSocketMessageType.Close)
+                {
+                    return ClientGone;
+                }
+
+                // A byte pipe has no text. A client sending some has the wrong idea about this endpoint,
+                // and forwarding it into an SSH stream would corrupt that stream at best.
+                if (received.MessageType != WebSocketMessageType.Binary)
+                {
+                    return new Ending(
+                        RelayCloseReason.Error,
+                        (int)WebSocketCloseStatus.InvalidMessageType,
+                        "The relay carries binary messages only.");
+                }
+
+                try
+                {
+                    await target.WriteAsync(buffer.AsMemory(0, received.Count), cancellationToken)
+                        .ConfigureAwait(false);
+                }
+                catch (OperationCanceledException)
+                {
+                    return null;
+                }
+                catch (IOException)
+                {
+                    return TargetGone;
+                }
+
+                counters.AddClientToTarget(received.Count);
+            }
+        }
+        finally
+        {
+            ArrayPool<byte>.Shared.Return(buffer);
+        }
+    }
+
+    /// <summary>Target to client: every read becomes one binary message.</summary>
+    private static async Task<Ending?> TargetToClientAsync(
+        Stream target,
+        WebSocket socket,
+        Counters counters,
+        CancellationToken cancellationToken)
+    {
+        var buffer = ArrayPool<byte>.Shared.Rent(BufferBytes);
+
+        try
+        {
+            while (true)
+            {
+                int read;
+
+                try
+                {
+                    read = await target.ReadAsync(buffer.AsMemory(0, BufferBytes), cancellationToken)
+                        .ConfigureAwait(false);
+                }
+                catch (OperationCanceledException)
+                {
+                    return null;
+                }
+                catch (IOException)
+                {
+                    return TargetGone;
+                }
+
+                if (read == 0)
+                {
+                    return TargetGone;
+                }
+
+                try
+                {
+                    // Not cancellable, deliberately. See PumpAsync.
+                    await socket
+                        .SendAsync(
+                            buffer.AsMemory(0, read),
+                            WebSocketMessageType.Binary,
+                            endOfMessage: true,
+                            CancellationToken.None)
+                        .ConfigureAwait(false);
+                }
+                catch (Exception exception) when (exception is WebSocketException or ObjectDisposedException)
+                {
+                    return ClientGone;
+                }
+
+                counters.AddTargetToClient(read);
+            }
+        }
+        finally
+        {
+            ArrayPool<byte>.Shared.Return(buffer);
+        }
+    }
+
+    /// <summary>
+    /// Ends the session at the idle timeout, the lifetime cap, or the end of a shutdown's drain.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// A shutdown does not end a session; it starts a countdown. ADR 0004 asks that a
+    /// <c>docker compose up -d</c> not guillotine live shells, so a session in progress when the server is
+    /// told to stop is given the drain timeout to finish — most will, because the user was already
+    /// closing it or the deployment waits — and is closed with 1001 only when that runs out, which tells
+    /// the client to reconnect rather than to report a fault. The host's own shutdown timeout is set
+    /// longer than the drain for this to be possible at all.
+    /// </para>
+    /// <para>
+    /// Idle means no bytes in either direction. WebSocket pings are not bytes of the stream and do not
+    /// count; SSH's own keepalives do, which is how a client that wants an idle shell kept open says so.
+    /// </para>
+    /// </remarks>
+    private async Task<Ending?> MindAsync(Counters counters, DateTimeOffset started, CancellationToken cancellationToken)
+    {
+        var settings = options.Value;
+        var hardDeadline = started + settings.MaxSessionDuration;
+        DateTimeOffset? drainDeadline = null;
+
+        while (!cancellationToken.IsCancellationRequested)
+        {
+            var now = clock.GetUtcNow();
+
+            if (drainDeadline is null && lifetime.ApplicationStopping.IsCancellationRequested)
+            {
+                drainDeadline = now + settings.DrainTimeout;
+            }
+
+            if (now >= hardDeadline)
+            {
+                return new Ending(
+                    RelayCloseReason.DurationLimit,
+                    RelayProtocol.DurationLimitCloseCode,
+                    "This relay session reached its maximum lifetime.");
+            }
+
+            if (drainDeadline is { } drain && now >= drain)
+            {
+                return ShuttingDown;
+            }
+
+            var idleDeadline = counters.LastActivity + settings.IdleTimeout;
+
+            if (now >= idleDeadline)
+            {
+                return new Ending(
+                    RelayCloseReason.IdleTimeout,
+                    RelayProtocol.IdleTimeoutCloseCode,
+                    "This relay session carried nothing for the idle timeout.");
+            }
+
+            var next = new[] { hardDeadline, idleDeadline, drainDeadline ?? DateTimeOffset.MaxValue }.Min();
+
+            // Woken early by the start of a shutdown, so the drain is measured from when it began rather
+            // than from whenever the idle clock next came round.
+            using var wake = drainDeadline is null
+                ? CancellationTokenSource.CreateLinkedTokenSource(cancellationToken, lifetime.ApplicationStopping)
+                : CancellationTokenSource.CreateLinkedTokenSource(cancellationToken);
+
+            try
+            {
+                await Task.Delay(next - now, clock, wake.Token).ConfigureAwait(false);
+            }
+            catch (OperationCanceledException) when (!cancellationToken.IsCancellationRequested)
+            {
+                // The shutdown began. Round again to start the drain.
+            }
+            catch (OperationCanceledException)
+            {
+                return null;
+            }
+        }
+
+        return null;
+    }
+
+    private static Ending ClientGone { get; } =
+        new(RelayCloseReason.ClientClosed, (int)WebSocketCloseStatus.NormalClosure, string.Empty);
+
+    private static Ending TargetGone { get; } =
+        new(RelayCloseReason.TargetClosed, (int)WebSocketCloseStatus.NormalClosure, "The host closed the connection.");
+
+    /// <summary>
+    /// Opens the audit row, before the target is dialled, so a session that never reached its host is
+    /// still one the operator can see.
+    /// </summary>
+    private async Task OpenAuditAsync(RelayTicket ticket, DateTimeOffset started)
+    {
+        var scope = scopes.CreateAsyncScope();
+        await using var _ = scope.ConfigureAwait(false);
+
+        var database = scope.ServiceProvider.GetRequiredService<DodoDbContext>();
+
+        database.RelaySessions.Add(new RelaySession
+        {
+            Id = ticket.SessionId,
+            UserId = ticket.UserId,
+            HostId = ticket.HostId,
+            VaultId = ticket.VaultId,
+            TargetHost = ticket.TargetHost,
+            TargetPort = ticket.TargetPort,
+            ClientAddress = HttpContext.Connection.RemoteIpAddress?.ToString(),
+            StartedAtUtc = started,
+        });
+
+        // Not the request's token: the row has to be written even for a client that has already gone,
+        // or a session that was cut off would be the one with no record.
+        await database.SaveChangesAsync(CancellationToken.None).ConfigureAwait(false);
+    }
+
+    /// <summary>
+    /// Closes the audit row. A failure is logged and swallowed: the bytes have already moved, and
+    /// throwing here would only turn a finished session into an unhandled exception.
+    /// </summary>
+    private async Task CloseAuditAsync(
+        Guid sessionId,
+        IPAddress? address,
+        Counters counters,
+        RelayCloseReason reason)
+    {
+        try
+        {
+            var scope = scopes.CreateAsyncScope();
+            await using var _ = scope.ConfigureAwait(false);
+
+            var database = scope.ServiceProvider.GetRequiredService<DodoDbContext>();
+            var ended = clock.GetUtcNow();
+            var dialled = address?.ToString();
+
+            await database.RelaySessions
+                .Where(session => session.Id == sessionId)
+                .ExecuteUpdateAsync(
+                    set => set
+                        .SetProperty(session => session.EndedAtUtc, ended)
+                        .SetProperty(session => session.TargetAddress, dialled)
+                        .SetProperty(session => session.BytesClientToTarget, counters.ClientToTarget)
+                        .SetProperty(session => session.BytesTargetToClient, counters.TargetToClient)
+                        .SetProperty(session => session.CloseReason, reason),
+                    CancellationToken.None)
+                .ConfigureAwait(false);
+        }
+        catch (Exception exception) when (exception is DbUpdateException or NpgsqlException or InvalidOperationException)
+        {
+            RelayLog.AuditCloseFailed(logger, sessionId, exception);
+        }
+    }
+
+    private static async Task CloseAsync(WebSocket socket, Ending ending)
+    {
+        if (socket.State is not (WebSocketState.Open or WebSocketState.CloseReceived))
+        {
+            return;
+        }
+
+        using var timeout = new CancellationTokenSource(CloseTimeout);
+
+        try
+        {
+            await socket
+                .CloseOutputAsync((WebSocketCloseStatus)ending.Code, ending.Description, timeout.Token)
+                .ConfigureAwait(false);
+        }
+        catch (Exception exception)
+            when (exception is OperationCanceledException or WebSocketException or ObjectDisposedException)
+        {
+            // The peer is already gone. There is nothing to tell it.
+        }
+    }
+
+    /// <summary>Waits for a loop to finish, for at most the close timeout.</summary>
+    /// <returns>Whether it finished.</returns>
+    private static async Task<bool> SettlesAsync(Task loop)
+    {
+        try
+        {
+            await loop.WaitAsync(CloseTimeout).ConfigureAwait(false);
+
+            return true;
+        }
+        catch (TimeoutException)
+        {
+            return false;
+        }
+        catch (Exception exception)
+            when (exception is OperationCanceledException or WebSocketException or IOException or ObjectDisposedException)
+        {
+            return true;
+        }
+    }
+
+    /// <summary>How a session ended, and what the client is told.</summary>
+    private sealed record Ending(RelayCloseReason Reason, int Code, string Description);
+
+    /// <summary>
+    /// Byte counts and the time of the last byte, written by the two directions and read by the clock.
+    /// </summary>
+    private sealed class Counters(TimeProvider clock, DateTimeOffset started)
+    {
+        private long clientToTarget;
+        private long targetToClient;
+        private long lastActivityTicks = started.UtcTicks;
+
+        internal long ClientToTarget => Interlocked.Read(ref clientToTarget);
+
+        internal long TargetToClient => Interlocked.Read(ref targetToClient);
+
+        internal DateTimeOffset LastActivity => new(Interlocked.Read(ref lastActivityTicks), TimeSpan.Zero);
+
+        internal void AddClientToTarget(int count)
+        {
+            Interlocked.Add(ref clientToTarget, count);
+            Interlocked.Exchange(ref lastActivityTicks, clock.GetUtcNow().UtcTicks);
+        }
+
+        internal void AddTargetToClient(int count)
+        {
+            Interlocked.Add(ref targetToClient, count);
+            Interlocked.Exchange(ref lastActivityTicks, clock.GetUtcNow().UtcTicks);
+        }
+    }
+}

src/DodoSSH.Api/Features/Relay/RelayLog.cs

@@ -0,0 +1,94 @@
+using DodoSSH.Domain;
+
+namespace DodoSSH.Api.Features.Relay;
+
+/// <summary>Source-generated log messages for the relay.</summary>
+/// <remarks>
+/// <para>
+/// Ids, counts and reasons. Never a ticket, which is a bearer credential for thirty seconds, and never
+/// the target's address: that is in the audit table, where retention is the operator's decision, and
+/// copying it into a log would make it the log pipeline's decision too.
+/// </para>
+/// <para>
+/// <see cref="TargetRejected"/> is the one place the reason for a refused target is written down. The
+/// client is told nothing specific, deliberately — see <c>RelayTicketEndpoint</c> — so this line is how
+/// an operator answers "why will the relay not reach this host".
+/// </para>
+/// </remarks>
+internal static partial class RelayLog
+{
+    [LoggerMessage(
+        EventId = 2301,
+        Level = LogLevel.Information,
+        Message = "Relay ticket refused for user {UserId} and host {HostId}: {Reason}.")]
+    internal static partial void TargetRejected(ILogger logger, Guid userId, Guid hostId, string reason);
+
+    [LoggerMessage(
+        EventId = 2302,
+        Level = LogLevel.Debug,
+        Message = "Relay ticket issued to user {UserId} for host {HostId}, session {SessionId}, with "
+            + "{AddressCount} permitted address(es).")]
+    internal static partial void TicketIssued(
+        ILogger logger,
+        Guid userId,
+        Guid hostId,
+        Guid sessionId,
+        int addressCount);
+
+    /// <remarks>
+    /// The session id is known only for a ticket that verified and had already been spent, which is the
+    /// case worth an operator's attention: a replay, or a client retrying a redemption it thinks failed.
+    /// </remarks>
+    [LoggerMessage(
+        EventId = 2303,
+        Level = LogLevel.Information,
+        Message = "Relay ticket refused at the socket (session {SessionId}): forged, expired or already used.")]
+    internal static partial void TicketRefused(ILogger logger, Guid? sessionId);
+
+    [LoggerMessage(
+        EventId = 2304,
+        Level = LogLevel.Information,
+        Message = "Relay session refused for user {UserId}: a session limit is met, with {SessionCount} "
+            + "open on this node.")]
+    internal static partial void LimitReached(ILogger logger, Guid userId, int sessionCount);
+
+    [LoggerMessage(
+        EventId = 2305,
+        Level = LogLevel.Information,
+        Message = "Relay session {SessionId} could not reach host {HostId} on any permitted address.")]
+    internal static partial void TargetUnreachable(ILogger logger, Guid sessionId, Guid hostId);
+
+    [LoggerMessage(
+        EventId = 2306,
+        Level = LogLevel.Debug,
+        Message = "Relay session {SessionId} opened for user {UserId} to host {HostId}; {SessionCount} "
+            + "open on this node.")]
+    internal static partial void SessionOpened(
+        ILogger logger,
+        Guid sessionId,
+        Guid userId,
+        Guid hostId,
+        int sessionCount);
+
+    [LoggerMessage(
+        EventId = 2307,
+        Level = LogLevel.Debug,
+        Message = "Relay session {SessionId} closed ({Reason}) after {ClientToTarget} byte(s) up and "
+            + "{TargetToClient} down.")]
+    internal static partial void SessionClosed(
+        ILogger logger,
+        Guid sessionId,
+        RelayCloseReason reason,
+        long clientToTarget,
+        long targetToClient);
+
+    /// <remarks>
+    /// A warning rather than an error: the session itself is over and was fine. What is lost is the
+    /// audit row's second half, which leaves it looking like a session still open.
+    /// </remarks>
+    [LoggerMessage(
+        EventId = 2308,
+        Level = LogLevel.Warning,
+        Message = "Relay session {SessionId} ended but its audit record could not be closed.")]
+    internal static partial void AuditCloseFailed(ILogger logger, Guid sessionId, Exception exception);
+}

src/DodoSSH.Api/Features/Relay/RelayNetwork.cs

@@ -0,0 +1,61 @@
+using System.Net;
+using System.Net.Sockets;
+
+namespace DodoSSH.Api.Features.Relay;
+
+/// <summary>
+/// The two things the relay asks of the network: what a name resolves to, and a connection to an address.
+/// </summary>
+/// <remarks>
+/// <para>
+/// A seam for tests, and a narrow one on purpose. <see cref="RelayAddressPolicy"/> is not behind it:
+/// whatever this answers, the policy judges, and a test that replaces this still runs every address
+/// through the real rules. What a test cannot do without the seam is dial anything, because every
+/// address a test machine can reach on its own — loopback — is one the policy refuses outright.
+/// </para>
+/// </remarks>
+internal interface IRelayNetwork
+{
+    /// <summary>Resolves a stored hostname, which may be an address literal.</summary>
+    Task<IPAddress[]> ResolveAsync(string host, CancellationToken cancellationToken);
+
+    /// <summary>Opens a TCP connection to one address.</summary>
+    /// <returns>A stream that owns the connection.</returns>
+    Task<Stream> ConnectAsync(IPAddress address, int port, CancellationToken cancellationToken);
+}
+
+/// <summary>The operating system's resolver and sockets.</summary>
+internal sealed class SystemRelayNetwork : IRelayNetwork
+{
+    /// <inheritdoc />
+    public Task<IPAddress[]> ResolveAsync(string host, CancellationToken cancellationToken) =>
+
+        // A literal parses to itself without a lookup, which is what an address typed into a host
+        // editor should do — and what makes "10.0.0.5" mean the same thing on every node.
+        IPAddress.TryParse(host, out var literal)
+            ? Task.FromResult<IPAddress[]>([literal])
+            : Dns.GetHostAddressesAsync(host, cancellationToken);
+
+    /// <inheritdoc />
+    public async Task<Stream> ConnectAsync(IPAddress address, int port, CancellationToken cancellationToken)
+    {
+        var socket = new Socket(address.AddressFamily, SocketType.Stream, ProtocolType.Tcp)
+        {
+            // An interactive shell is a stream of keystrokes; Nagle would hold each one back for the
+            // acknowledgement of the last, which is the latency the user feels as a sticky terminal.
+            NoDelay = true,
+        };
+
+        try
+        {
+            await socket.ConnectAsync(new IPEndPoint(address, port), cancellationToken).ConfigureAwait(false);
+
+            return new NetworkStream(socket, ownsSocket: true);
+        }
+        catch
+        {
+            socket.Dispose();
+            throw;
+        }
+    }
+}

src/DodoSSH.Api/Features/Relay/RelaySessionRegistry.cs

@@ -0,0 +1,111 @@
+using DodoSSH.Api.Setup;
+using Microsoft.Extensions.Options;
+
+namespace DodoSSH.Api.Features.Relay;
+
+/// <summary>
+/// Counts the relay sessions open on this node, and refuses the ones over ADR 0004's limits.
+/// </summary>
+/// <remarks>
+/// <para>
+/// Per node, and so per user <em>per node</em>: a user behind a round-robin balancer with three nodes
+/// could hold three times the per-user cap. That is accepted rather than fixed, because fixing it means
+/// shared state on the hot path of every connect, and the cap exists to stop one account exhausting a
+/// node — which a per-node count does exactly.
+/// </para>
+/// <para>
+/// A singleton, because a session outlives the request that opened it. It holds counts and nothing else;
+/// the sockets belong to the handlers that opened them.
+/// </para>
+/// </remarks>
+internal sealed class RelaySessionRegistry(IOptions<RelayOptions> options)
+{
+    private readonly Lock gate = new();
+    private readonly Dictionary<Guid, int> perUser = [];
+    private int total;
+
+    /// <summary>How many sessions are open on this node.</summary>
+    internal int Count
+    {
+        get
+        {
+            lock (gate)
+            {
+                return total;
+            }
+        }
+    }
+
+    /// <summary>Whether a session for this user would be admitted now.</summary>
+    /// <remarks>
+    /// Advisory, for the ticket endpoint: it lets a caller already at the cap hear so as an HTTP status
+    /// before it has spent a ticket, but only <see cref="TryAdmit"/> decides, since the answer here can
+    /// be stale by the time the socket arrives — possibly at another node.
+    /// </remarks>
+    internal bool HasRoomFor(Guid userId)
+    {
+        var settings = options.Value;
+
+        lock (gate)
+        {
+            return total < settings.MaxConcurrentSessionsTotal
+                && perUser.GetValueOrDefault(userId) < settings.MaxConcurrentSessionsPerUser;
+        }
+    }
+
+    /// <summary>Admits a session, or answers null when a limit is met.</summary>
+    /// <returns>A lease that gives the place back when disposed.</returns>
+    internal Lease? TryAdmit(Guid userId)
+    {
+        var settings = options.Value;
+
+        lock (gate)
+        {
+            var held = perUser.GetValueOrDefault(userId);
+
+            if (total >= settings.MaxConcurrentSessionsTotal || held >= settings.MaxConcurrentSessionsPerUser)
+            {
+                return null;
+            }
+
+            perUser[userId] = held + 1;
+            total++;
+
+            return new Lease(this, userId);
+        }
+    }
+
+    private void Release(Guid userId)
+    {
+        lock (gate)
+        {
+            total--;
+
+            // Removed at zero rather than left behind, so the dictionary is the size of the users with a
+            // session open and not of every user who ever had one.
+            if (perUser[userId] <= 1)
+            {
+                perUser.Remove(userId);
+            }
+            else
+            {
+                perUser[userId]--;
+            }
+        }
+    }
+
+    /// <summary>One admitted session's place. Idempotent to dispose.</summary>
+    internal sealed class Lease(RelaySessionRegistry registry, Guid userId) : IDisposable
+    {
+        private int released;
+
+        /// <inheritdoc />
+        public void Dispose()
+        {
+            if (Interlocked.Exchange(ref released, 1) == 0)
+            {
+                registry.Release(userId);
+            }
+        }
+    }
+}

src/DodoSSH.Api/Features/Relay/RelayTicketCodec.cs

@@ -0,0 +1,171 @@
+using System.Net;
+using System.Security.Cryptography;
+using Microsoft.AspNetCore.DataProtection;
+using Microsoft.AspNetCore.WebUtilities;
+
+namespace DodoSSH.Api.Features.Relay;
+
+/// <summary>
+/// Everything the relay socket needs to know, decided by the ticket endpoint.
+/// </summary>
+/// <remarks>
+/// The addresses are the resolved ones and the socket dials nothing else. There is no hostname for it
+/// to resolve again — <see cref="TargetHost"/> is carried for the audit row only — which is what keeps a
+/// DNS answer that changes between issue and connect from changing where the bytes go.
+/// </remarks>
+/// <param name="TicketId">The replay guard's key.</param>
+/// <param name="SessionId">The audit row's id, already given to the client.</param>
+/// <param name="UserId">Who asked.</param>
+/// <param name="HostId">The host it is for.</param>
+/// <param name="VaultId">The host's vault at issue.</param>
+/// <param name="TargetHost">The hostname as stored, for the audit row.</param>
+/// <param name="TargetAddresses">The permitted addresses the name resolved to, in the order to try.</param>
+/// <param name="TargetPort">The port.</param>
+/// <param name="ExpiresAt">When the ticket stops verifying.</param>
+internal sealed record RelayTicket(
+    Guid TicketId,
+    Guid SessionId,
+    Guid UserId,
+    Guid HostId,
+    Guid VaultId,
+    string TargetHost,
+    IReadOnlyList<IPAddress> TargetAddresses,
+    int TargetPort,
+    DateTimeOffset ExpiresAt);
+
+/// <summary>
+/// Seals and opens relay tickets.
+/// </summary>
+/// <remarks>
+/// <para>
+/// A Data Protection payload, per ADR 0004: authenticated and encrypted, with key rotation handled by
+/// the framework, and verifiable on any node that shares the key ring. Encrypted rather than merely
+/// signed because the ticket travels in a header every proxy on the way can read, and what it holds is
+/// the resolved address of a machine on the operator's network.
+/// </para>
+/// <para>
+/// The expiry is inside the payload and checked against <see cref="TimeProvider"/>, rather than left to
+/// the framework's time-limited protector, which reads the system clock directly — the thing this
+/// repository bans so that time can be faked in tests.
+/// </para>
+/// <para>
+/// The layout is a version byte and fixed fields, written by hand. Small enough that a serialiser would
+/// be the larger half of the code, and a hand layout has no reflection for trimming to break.
+/// </para>
+/// </remarks>
+internal sealed class RelayTicketCodec(IDataProtectionProvider protection, TimeProvider clock)
+{
+    /// <summary>The purpose string, which isolates these payloads from every other protector's.</summary>
+    private const string Purpose = "DodoSSH.Relay.Ticket";
+
+    private const byte LayoutVersion = 1;
+
+    /// <summary>More addresses than any real name resolves to, and a bound on what a ticket can carry.</summary>
+    internal const int MaxAddresses = 16;
+
+    private readonly IDataProtector protector = protection.CreateProtector(Purpose);
+
+    internal string Seal(RelayTicket ticket)
+    {
+        ArgumentNullException.ThrowIfNull(ticket);
+
+        using var buffer = new MemoryStream();
+        using (var writer = new BinaryWriter(buffer))
+        {
+            writer.Write(LayoutVersion);
+            writer.Write(ticket.TicketId.ToByteArray());
+            writer.Write(ticket.SessionId.ToByteArray());
+            writer.Write(ticket.UserId.ToByteArray());
+            writer.Write(ticket.HostId.ToByteArray());
+            writer.Write(ticket.VaultId.ToByteArray());
+            writer.Write(ticket.ExpiresAt.ToUnixTimeMilliseconds());
+            writer.Write(ticket.TargetPort);
+            writer.Write(ticket.TargetHost);
+            writer.Write((byte)ticket.TargetAddresses.Count);
+
+            foreach (var address in ticket.TargetAddresses)
+            {
+                var bytes = address.GetAddressBytes();
+                writer.Write((byte)bytes.Length);
+                writer.Write(bytes);
+            }
+        }
+
+        // Base64url, whose alphabet is all token characters, so it can stand in a subprotocol name.
+        return WebEncoders.Base64UrlEncode(protector.Protect(buffer.ToArray()));
+    }
+
+    /// <summary>Opens a ticket, or answers null for one that is forged, damaged or expired.</summary>
+    /// <remarks>
+    /// One answer for all three, because the caller's remedy is the same — ask for another — and a
+    /// distinct "expired" would tell a forger which of their attempts had verified.
+    /// </remarks>
+    internal RelayTicket? TryOpen(string sealedTicket)
+    {
+        byte[] payload;
+
+        try
+        {
+            payload = protector.Unprotect(WebEncoders.Base64UrlDecode(sealedTicket));
+        }
+        catch (Exception exception) when (exception is CryptographicException or FormatException)
+        {
+            return null;
+        }
+
+        RelayTicket ticket;
+
+        try
+        {
+            ticket = Read(payload);
+        }
+        catch (Exception exception)
+            when (exception is EndOfStreamException or ArgumentException or FormatException)
+        {
+            return null;
+        }
+
+        return ticket.ExpiresAt > clock.GetUtcNow() ? ticket : null;
+    }
+
+    private static RelayTicket Read(byte[] payload)
+    {
+        using var reader = new BinaryReader(new MemoryStream(payload));
+
+        if (reader.ReadByte() != LayoutVersion)
+        {
+            throw new FormatException("Unknown ticket layout.");
+        }
+
+        var ticketId = new Guid(reader.ReadBytes(16));
+        var sessionId = new Guid(reader.ReadBytes(16));
+        var userId = new Guid(reader.ReadBytes(16));
+        var hostId = new Guid(reader.ReadBytes(16));
+        var vaultId = new Guid(reader.ReadBytes(16));
+        var expiresAt = DateTimeOffset.FromUnixTimeMilliseconds(reader.ReadInt64());
+        var port = reader.ReadInt32();
+        var host = reader.ReadString();
+        var count = reader.ReadByte();
+
+        if (count is 0 or > MaxAddresses)
+        {
+            throw new FormatException("A ticket names between one and sixteen addresses.");
+        }
+
+        var addresses = new IPAddress[count];
+
+        for (var index = 0; index < count; index++)
+        {
+            var length = reader.ReadByte();
+
+            if (length is not (4 or 16))
+            {
+                throw new FormatException("An address is four or sixteen bytes.");
+            }
+
+            addresses[index] = new IPAddress(reader.ReadBytes(length));
+        }
+
+        return new RelayTicket(ticketId, sessionId, userId, hostId, vaultId, host, addresses, port, expiresAt);
+    }
+}

src/DodoSSH.Api/Features/Relay/RelayTicketEndpoint.cs

@@ -0,0 +1,231 @@
+using System.Net;
+using System.Net.Sockets;
+using DodoSSH.Api.Authorization;
+using DodoSSH.Api.Setup;
+using DodoSSH.Contracts;
+using DodoSSH.Domain.Authorization;
+using DodoSSH.Infrastructure;
+using FastEndpoints;
+using Microsoft.AspNetCore.Http.HttpResults;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.Options;
+
+namespace DodoSSH.Api.Features.Relay;
+
+/// <summary>
+/// Issues the ticket that lets exactly one relay socket reach exactly one host.
+/// </summary>
+/// <remarks>
+/// <para>
+/// Every decision the relay makes is made here, with the full bearer token and the full access model in
+/// hand: which host, whether the caller may connect to it, which addresses its name resolves to and
+/// which of those the policy permits. The socket that follows decides nothing. It checks that the ticket
+/// is genuine, fresh and unused, and dials what the ticket says. See ADR 0004.
+/// </para>
+/// <para>
+/// <b>Every refusal about the target is the same refusal.</b> A host that does not exist, one in a vault
+/// the caller cannot reach, one without the relay switched on, one whose name does not resolve, and one
+/// whose every address the policy refuses all answer 403 <c>relay-target-rejected</c> with the same
+/// sentence. Anyone who can write a host can set its address to anything, so a distinct answer for
+/// "resolves, but refused" would turn this endpoint into a scanner for the operator's internal DNS. The
+/// real reason goes to the server log, where the operator can read it.
+/// </para>
+/// </remarks>
+internal sealed class RelayTicketEndpoint(
+    ICurrentUserContext currentUser,
+    IVaultAccessService vaultAccess,
+    DodoDbContext database,
+    IRelayNetwork network,
+    RelayTicketCodec codec,
+    RelaySessionRegistry sessions,
+    IOptions<RelayOptions> options,
+    TimeProvider clock,
+    ILogger<RelayTicketEndpoint> logger)
+    : Endpoint<RelayTicketRequest, Results<Ok<RelayTicketResponse>, ProblemHttpResult>>
+{
+    /// <summary>What every target refusal says, whatever the reason was.</summary>
+    private const string RejectedDetail =
+        "The relay will not connect to this host. Check that the relay is switched on for it and that its "
+        + "address is one this server may reach; the server's log records which.";
+
+    /// <inheritdoc />
+    public override void Configure()
+    {
+        Post(RelayProtocol.TicketsPath);
+
+        // Enrolled, matching sync. A relayed host is still a host whose name, username and key are
+        // ciphertext, so a caller with no identity key has no way to have learned which host to ask for.
+        Policies(Auth.EnrolledPolicy);
+
+        Description(b => b
+            .WithName("IssueRelayTicket")
+            .WithSummary("Issues a single-use ticket for one relay connection to one host.")
+            .WithTags("Relay"));
+    }
+
+    /// <inheritdoc />
+    public override async Task<Results<Ok<RelayTicketResponse>, ProblemHttpResult>> ExecuteAsync(
+        RelayTicketRequest req,
+        CancellationToken ct)
+    {
+        var settings = options.Value;
+
+        if (!settings.Enabled)
+        {
+            return Problems.Coded(
+                StatusCodes.Status404NotFound,
+                ProblemCodes.RelayUnavailable,
+                "This server does not relay. Connect to the host directly; GET /api/v1/meta lists the "
+                + "features this server does offer.");
+        }
+
+        var user = await currentUser.GetOrProvisionAsync(ct).ConfigureAwait(false);
+        var (target, refusal) = await FindTargetAsync(user.Id, req, ct).ConfigureAwait(false);
+
+        if (target is not var (hostname, port, vaultId))
+        {
+            return Rejected(user.Id, req.HostId, refusal!);
+        }
+
+        // Before resolving rather than after: a caller already at the cap is told so without this server
+        // spending a DNS query on a ticket that could not be redeemed.
+        if (!sessions.HasRoomFor(user.Id))
+        {
+            RelayLog.LimitReached(logger, user.Id, sessions.Count);
+
+            return Problems.Coded(
+                StatusCodes.Status429TooManyRequests,
+                ProblemCodes.RelayLimitReached,
+                "You have as many relay sessions open as this server allows. Close one and try again.");
+        }
+
+        var addresses = await ResolvePermittedAsync(hostname, ct).ConfigureAwait(false);
+
+        if (addresses is not { Length: > 0 })
+        {
+            return Rejected(user.Id, req.HostId, addresses is null
+                ? "its name did not resolve"
+                : "every address its name resolved to is refused by policy");
+        }
+
+        var ticket = new RelayTicket(
+            TicketId: Guid.CreateVersion7(),
+            SessionId: Guid.CreateVersion7(),
+            UserId: user.Id,
+            HostId: req.HostId,
+            VaultId: vaultId,
+            TargetHost: hostname,
+            TargetAddresses: addresses,
+            TargetPort: port,
+            ExpiresAt: clock.GetUtcNow() + settings.TicketLifetime);
+
+        RelayLog.TicketIssued(logger, user.Id, req.HostId, ticket.SessionId, addresses.Length);
+
+        return TypedResults.Ok(new RelayTicketResponse(
+            Ticket: codec.Seal(ticket),
+            WebSocketUrl: new Uri(settings.WebSocketUrl!, UriKind.Absolute),
+            SubProtocol: RelayProtocol.SubProtocol,
+            SessionId: ticket.SessionId,
+            ExpiresAt: ticket.ExpiresAt));
+    }
+
+    /// <summary>
+    /// Finds the host, and decides whether this caller may be relayed to it at all.
+    /// </summary>
+    /// <returns>The stored target, or the reason there is none — for the log, never for the caller.</returns>
+    private async Task<((string Hostname, int Port, Guid VaultId)? Target, string? Refusal)> FindTargetAsync(
+        Guid userId,
+        RelayTicketRequest req,
+        CancellationToken cancellationToken)
+    {
+        if (req.PortForwardId is not null)
+        {
+            return (null, "a port forward was requested, which v1 does not relay");
+        }
+
+        var host = await database.Hosts
+            .AsNoTracking()
+            .SingleOrDefaultAsync(h => h.Id == req.HostId && h.DeletedAtUtc == null, cancellationToken)
+            .ConfigureAwait(false);
+
+        if (host is null)
+        {
+            return (null, "no such host");
+        }
+
+        var access = await vaultAccess.ResolveAsync(userId, host.VaultId, cancellationToken).ConfigureAwait(false);
+
+        // Connect rather than Read, because Connect is the permission that says what this is. Every role
+        // that reads today also connects — see VaultAccessService.ForRole — but a role that ever split
+        // them should be refused a relay here without anybody remembering to look.
+        if (!access.Granted || !access.Permissions.HasFlag(PermissionFlags.Connect))
+        {
+            return (null, "the caller may not connect to this host's vault");
+        }
+
+        // The CHECK constraint makes the second half unreachable while the first is true. Tested anyway,
+        // because the cost of trusting it is dialling a null.
+        if (!host.RelayEnabled || host.Hostname is not { } hostname || host.Port is not { } port)
+        {
+            return (null, "the relay is not switched on for this host");
+        }
+
+        return RelayAddressPolicy.PermitsPort(port)
+            ? ((hostname, port, host.VaultId), null)
+            : (null, $"port {port} is refused by policy");
+    }
+
+    /// <summary>
+    /// Resolves the stored name, and keeps the addresses the policy permits.
+    /// </summary>
+    /// <returns>The permitted addresses, possibly none; or null when the name did not resolve.</returns>
+    /// <remarks>
+    /// <para>
+    /// Filtered rather than refused as a whole when one answer is bad. A name that resolves to a public
+    /// address and a link-local one is dialled on the public address, and never on the other — the
+    /// ticket carries only what passed, so there is nothing else for the socket to try.
+    /// </para>
+    /// <para>
+    /// The addresses kept are the normalised ones, so what is dialled is exactly what was judged rather
+    /// than a second spelling of it.
+    /// </para>
+    /// </remarks>
+    private async Task<IPAddress[]?> ResolvePermittedAsync(string hostname, CancellationToken cancellationToken)
+    {
+        var policy = new RelayAddressPolicy(options.Value);
+
+        using var timeout = CancellationTokenSource.CreateLinkedTokenSource(cancellationToken);
+        timeout.CancelAfter(options.Value.ConnectTimeout);
+
+        IPAddress[] resolved;
+
+        try
+        {
+            resolved = await network.ResolveAsync(hostname, timeout.Token).ConfigureAwait(false);
+        }
+        catch (SocketException)
+        {
+            return null;
+        }
+        catch (OperationCanceledException) when (!cancellationToken.IsCancellationRequested)
+        {
+            return null;
+        }
+
+        return
+        [
+            .. resolved
+                .Where(policy.Permits)
+                .Select(RelayAddressPolicy.Normalise)
+                .Distinct()
+                .Take(RelayTicketCodec.MaxAddresses),
+        ];
+    }
+
+    private ProblemHttpResult Rejected(Guid userId, Guid hostId, string reason)
+    {
+        RelayLog.TargetRejected(logger, userId, hostId, reason);
+
+        return Problems.Coded(StatusCodes.Status403Forbidden, ProblemCodes.RelayTargetRejected, RejectedDetail);
+    }
+}

src/DodoSSH.Api/Program.cs

@@ -49,6 +49,9 @@ builder.Services.AddSingleton<VaultEventHub>();
 builder.Services.AddSingleton<IVaultEventPublisher>(
     provider => provider.GetRequiredService<VaultEventHub>());
 
+// The relay's services, its ticket key ring, and a shutdown long enough for its drain. See ADR 0004.
+builder.Services.AddDodoRelay();
+
 // Scoped rather than the AddAuthorization default of singleton: the handler reads the request's
 // DbContext, and a singleton would capture one for the lifetime of the process.
 builder.Services.AddScoped<IAuthorizationHandler, EnrolledHandler>();

src/DodoSSH.Api/Setup/Configuration.cs

@@ -54,6 +54,13 @@ internal static class Configuration
                 + "setting exists to provide.")
             .ValidateOnStart();
 
+        services.AddRelayOptions();
+
+        return services;
+    }
+
+    /// <summary>The relay's options, which carry more rules than any other section. See ADR 0004.</summary>
+    private static void AddRelayOptions(this IServiceCollection services) =>
         services.AddOptions<RelayOptions>()
             .BindConfiguration(RelayOptions.SectionName)
             .ValidateDataAnnotations()
@@ -63,11 +70,28 @@ internal static class Configuration
             .Validate(
                 options => options.IdleTimeout < options.MaxSessionDuration,
                 "Relay:IdleTimeout must be shorter than Relay:MaxSessionDuration.")
+            .Validate(
+                options => options.TicketLifetime > TimeSpan.Zero
+                    && options.TicketLifetime <= TimeSpan.FromSeconds(30),
+                "Relay:TicketLifetime must be positive and at most 30 seconds. A ticket is a bearer "
+                + "credential in a header; ADR 0004 bounds how long one may be worth stealing.")
+            .Validate(
+                options => options.ConnectTimeout > TimeSpan.Zero && options.DrainTimeout >= TimeSpan.Zero,
+                "Relay:ConnectTimeout must be positive and Relay:DrainTimeout must not be negative.")
+            .Validate(
+                options => Features.Relay.RelayAddressPolicy.AreNetworks(options.AllowedNetworks)
+                    && Features.Relay.RelayAddressPolicy.AreNetworks(options.DeniedNetworks),
+                "Relay:AllowedNetworks and Relay:DeniedNetworks must hold CIDR networks such as "
+                + "10.20.0.0/16, with no host bits set. An entry this server cannot read would be a gap "
+                + "the operator believes is closed.")
+            .Validate(
+                options => !options.Enabled
+                    || (Uri.TryCreate(options.WebSocketUrl, UriKind.Absolute, out var url)
+                        && (string.Equals(url.Scheme, Uri.UriSchemeWs, StringComparison.Ordinal)
+                            || string.Equals(url.Scheme, Uri.UriSchemeWss, StringComparison.Ordinal))),
+                "Relay:WebSocketUrl must be an absolute ws:// or wss:// URL.")
             .ValidateOnStart();
 
-        return services;
-    }
-
     /// <summary>
     /// Warns about configuration that is individually valid but dangerous in combination.
     /// </summary>

src/DodoSSH.Api/Setup/DodoOptions.cs

@@ -115,13 +115,50 @@ public sealed class RelayOptions
     /// Whether RFC1918 and other private ranges may be dialled.
     /// </summary>
     /// <remarks>
+    /// <para>
     /// Defaults to true, unlike a typical SSRF allow-list, because reaching private
     /// infrastructure is the entire purpose of a self-hosted SSH tool. The control that matters is
-    /// the ACL: only hosts in a vault the caller holds Connect on can be dialled at all. Loopback,
-    /// link-local and cloud metadata ranges are denied unconditionally and are not configurable.
+    /// the ACL: only hosts in a vault the caller holds Connect on can be dialled at all.
+    /// </para>
+    /// <para>
+    /// Governs the private ranges only — RFC1918, carrier-grade NAT (which is where a Tailnet lives) and
+    /// IPv6 unique-local — and never the list <c>RelayAddressPolicy</c> refuses with no configuration
+    /// escape: loopback, link-local, <c>0.0.0.0/8</c>, multicast and the cloud metadata addresses.
+    /// Nothing here can re-open those, which is what ADR 0004 §3 means by non-overridable.
+    /// </para>
     /// </remarks>
     public bool AllowPrivateNetworks { get; set; } = true;
 
+    /// <summary>
+    /// Networks, in CIDR form, that a relay target must fall inside. Empty means any network the other
+    /// rules permit.
+    /// </summary>
+    /// <remarks>
+    /// When set, this is the whole answer for the ranges it names: a target inside one is permitted even
+    /// with <see cref="AllowPrivateNetworks"/> off, because "only 10.20.0.0/16" is the precise thing an
+    /// operator means by setting both. It still cannot reach the non-overridable list, and
+    /// <see cref="DeniedNetworks"/> still wins over it.
+    /// </remarks>
+    public IList<string> AllowedNetworks { get; } = [];
+
+    /// <summary>Networks, in CIDR form, that a relay target may never fall inside.</summary>
+    /// <remarks>
+    /// Checked before <see cref="AllowedNetworks"/>, so a deny inside an allow is a hole in it — the way
+    /// round that lets an operator carve the database subnet out of an otherwise permitted range.
+    /// </remarks>
+    public IList<string> DeniedNetworks { get; } = [];
+
+    /// <summary>
+    /// A directory to keep the ticket key ring in. Unset means the platform default.
+    /// </summary>
+    /// <remarks>
+    /// Losing the key ring is harmless in the way losing <c>Sync:CursorSigningKey</c> is: a ticket lives
+    /// thirty seconds, so a restart invalidates at most the connects in flight and the client asks
+    /// again. What does need it is more than one node — every node must be able to read every other's
+    /// tickets for round-robin to work, so a multi-node deployment points this at shared storage.
+    /// </remarks>
+    public string? KeyRingDirectory { get; set; }
+
     /// <summary>Maximum concurrent sessions per user.</summary>
     [Range(1, 1000)]
     public int MaxConcurrentSessionsPerUser { get; set; } = 10;

src/DodoSSH.Api/Setup/EndpointRegistration.cs

@@ -1,6 +1,7 @@
 using DodoSSH.Api.Features.Events;
 using DodoSSH.Api.Features.Identity;
 using DodoSSH.Api.Features.Meta;
+using DodoSSH.Api.Features.Relay;
 using DodoSSH.Api.Features.Sync;
 using DodoSSH.Api.Features.Teams;
 using DodoSSH.Contracts;
@@ -45,6 +46,8 @@ internal static class EndpointRegistration
             typeof(SyncPullEndpoint),
             typeof(SyncPushEndpoint),
             typeof(VaultEventsEndpoint),
+            typeof(RelayTicketEndpoint),
+            typeof(RelayConnectEndpoint),
             typeof(CreateTeamEndpoint),
             typeof(ListTeamsEndpoint),
             typeof(UpdateTeamEndpoint),
@@ -65,7 +68,6 @@ internal static class EndpointRegistration
             // Registered as each feature lands:
             //   Identity   — key rotation, passphrase change
             //   Vaults     — per-item ACLs
-            //   Relay      — tickets and the WebSocket
             //   Audit, Admin
         });
 

src/DodoSSH.Api/Setup/RelaySetup.cs

@@ -0,0 +1,59 @@
+using DodoSSH.Api.Features.Relay;
+using Microsoft.AspNetCore.DataProtection;
+using Microsoft.AspNetCore.DataProtection.KeyManagement;
+using Microsoft.AspNetCore.DataProtection.Repositories;
+using Microsoft.Extensions.Options;
+
+namespace DodoSSH.Api.Setup;
+
+/// <summary>The relay's services. See ADR 0004.</summary>
+internal static class RelaySetup
+{
+    /// <summary>
+    /// How much longer than the drain the host waits before it stops regardless.
+    /// </summary>
+    /// <remarks>
+    /// Room for the close handshakes and the audit rows' final writes once the drain has run out, which
+    /// is the part of a shutdown an operator would otherwise lose without noticing.
+    /// </remarks>
+    private static readonly TimeSpan ShutdownMargin = TimeSpan.FromSeconds(10);
+
+    internal static IServiceCollection AddDodoRelay(this IServiceCollection services)
+    {
+        // Singletons: a session outlives the request that opened it, a protector is safe to share, and
+        // the network seam holds no state.
+        services.AddSingleton<RelayTicketCodec>();
+        services.AddSingleton<RelaySessionRegistry>();
+        services.AddSingleton<IRelayNetwork, SystemRelayNetwork>();
+
+        // The application name isolates this deployment's payloads from any other ASP.NET application
+        // sharing a key directory, and — unlike the default, which is derived from the content root —
+        // is the same on every node, which is what lets any node verify any node's ticket.
+        services.AddDataProtection().SetApplicationName("DodoSSH");
+
+        services.AddOptions<KeyManagementOptions>()
+            .Configure<IOptions<RelayOptions>, ILoggerFactory>((keys, relay, loggers) =>
+            {
+                if (relay.Value.KeyRingDirectory is { Length: > 0 } directory)
+                {
+                    keys.XmlRepository = new FileSystemXmlRepository(new DirectoryInfo(directory), loggers);
+                }
+            });
+
+        // Longer than the drain, or the host would stop waiting for live sessions before the drain ADR 0004
+        // promises them had run. Harmless with the relay off: nothing else holds a request open that long,
+        // and the host stops as soon as the last request has.
+        services.AddOptions<HostOptions>()
+            .Configure<IOptions<RelayOptions>>((host, relay) =>
+            {
+                var wanted = relay.Value.DrainTimeout + ShutdownMargin;
+
+                if (wanted > host.ShutdownTimeout)
+                {
+                    host.ShutdownTimeout = wanted;
+                }
+            });
+
+        return services;
+    }
+}

src/DodoSSH.Client.Android/App.axaml.cs

@@ -86,36 +86,71 @@ public sealed partial class DodoSshApp : Avalonia.Application
         var shell = new PhoneShell();
 
         var viewModel = ComposeShell(
-            paths, caches, workspace, knownHosts, connections, ClipboardWriter(shell));
+            paths,
+            caches,
+            workspace,
+            knownHosts,
+            connections,
+            ClipboardWriter(shell),
+            ClipboardReader(shell));
 
         shell.DataContext = viewModel;
 
-        // Difference 2: the foreground service, which is what makes TerminalWorkspace's promise — that a
-        // shell outlives a vault lock — true on a platform that stops backgrounded processes.
-        //
-        // The transfer count is real now that the document picker gives this head a way to start one, and
-        // it is the half that matters most here: a shell survives backgrounding because somebody is looking
-        // at it, and an upload has to survive precisely when nobody is — the screen is off and the phone is
-        // in a pocket. Queued counts as active, so putting five files in the queue and locking the phone
-        // moves five files.
-        //
-        // A local rather than a field, matching the desktop head: an Avalonia Application has no disposal
-        // hook, so a field holding a disposable would have nowhere honest to release it. It stays alive
-        // because it is subscribed to the workspace, which lives as long as the process.
-        var keepAlive = new SessionKeepAlive(
-            workspace,
-            activeTransfers: () => viewModel.Transfers.ActiveTransfers);
-
-        // The other end of the same wire: the workspace announces its own sessions ending, and the queue
-        // announces transfers appearing and finishing. Without this the notification would come up when an
-        // upload started and stay up after it finished, which is the failure this class exists to prevent.
-        viewModel.Transfers.ActivityChanged += (_, _) => keepAlive.Refresh();
-
-        keepAlive.Refresh();
+        // Difference 2, wired up in its own method purely for length — see ComposeKeepAlive for what it
+        // does and why.
+        ComposeKeepAlive(workspace, viewModel);
 
         return shell;
     }
 
+    /// <summary>
+    /// Wires up the foreground service that makes <c>TerminalWorkspace</c>'s promise — that a shell outlives
+    /// a vault lock — true on a platform that stops backgrounded processes.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// Split out of <see cref="Compose"/> for length rather than for reuse; there is exactly one caller.
+    /// </para>
+    /// <para>
+    /// The transfer count is real now that the document picker gives this head a way to start one, and it
+    /// matters exactly when nobody is looking: a shell survives backgrounding because somebody opened it,
+    /// and an upload has to survive precisely when nobody is — the screen is off and the phone is in a
+    /// pocket. Queued counts as active, so putting five files in the queue and locking the phone moves five
+    /// files. <c>holdsFileSession</c> covers the third case a count alone cannot: a host connected on the
+    /// Files screen with no transfer moving is still a live SFTP session that backgrounding would sever, and
+    /// <c>TransfersViewModel.HasLiveFileSession</c> is the existing fact — <c>IsConnected</c> with a real
+    /// cipher, which a bucket never has — that answers whether one is open.
+    /// </para>
+    /// <para>
+    /// <c>keepAlive</c> is a local rather than a field, matching the desktop head: an Avalonia Application
+    /// has no disposal hook, so a field holding a disposable would have nowhere honest to release it. It
+    /// stays alive because it is subscribed to the workspace, which lives as long as the process.
+    /// </para>
+    /// </remarks>
+    private static void ComposeKeepAlive(TerminalWorkspace workspace, MainWindowViewModel viewModel)
+    {
+        var keepAlive = new SessionKeepAlive(
+            workspace,
+            activeTransfers: () => viewModel.Transfers.ActiveTransfers,
+            holdsFileSession: () => viewModel.Transfers.HasLiveFileSession);
+
+        // The other end of the same wire: the workspace announces its own sessions ending, and the queue
+        // announces transfers appearing and finishing, and a Files session connecting or disconnecting.
+        // Without this the notification would come up when an upload started and stay up after it
+        // finished, which is the failure this class exists to prevent.
+        viewModel.Transfers.ActivityChanged += (_, _) => keepAlive.Refresh();
+
+        // The half that was missing until now: a shell opening. SessionKeepAlive already heard the
+        // workspace announce a session ending, but nothing announced the opposite — a user who opened a
+        // shell and backgrounded the app had no foreground service at all, because the only wire in was the
+        // one for taking it down. TerminalSessionOpened is that other half, forwarded from
+        // VaultViewModel.SessionOpened, and without this line the service could never come up for a shell
+        // in the first place, which was precisely the promise this whole arrangement exists to keep.
+        viewModel.TerminalSessionOpened += (_, _) => keepAlive.Refresh();
+
+        keepAlive.Refresh();
+    }
+
     /// <summary>
     /// Writing to this phone's clipboard.
     /// </summary>
@@ -147,6 +182,26 @@ public sealed partial class DodoSshApp : Avalonia.Application
             }
         };
 
+    /// <summary>Reads this phone's clipboard.</summary>
+    /// <remarks>
+    /// <para>
+    /// The other direction, looked up per call for the same reason and with the same fall-through. One
+    /// caller: the terminal's right-click paste, which the renderer cannot serve itself — see
+    /// <c>TerminalClientOpcode.ClipboardPaste</c>.
+    /// </para>
+    /// <para>
+    /// Wired on this head although a phone has no right button, and not for symmetry: this head runs the same
+    /// renderer page, and a device with a mouse, a trackpad case or a desktop dock has a secondary click.
+    /// Whether Android's WebView delivers one as a DOM <c>contextmenu</c> event is unverified — see
+    /// <c>docs/manual-checks.md</c> 11.13 — and leaving this out in the meantime would make the paste half
+    /// the one thing that worked on one head and silently did nothing on the other.
+    /// </para>
+    /// </remarks>
+    private static Func<Task<string?>> ClipboardReader(Visual surface) =>
+        async () => TopLevel.GetTopLevel(surface) is { Clipboard: { } clipboard }
+            ? await clipboard.TryGetTextAsync().ConfigureAwait(false)
+            : null;
+
     /// <remarks>
     /// Split from <see cref="Compose"/> only for length. The division is a real one though: above this is
     /// the platform graph, and below it is the shell every head shares.
@@ -157,7 +212,8 @@ public sealed partial class DodoSshApp : Avalonia.Application
         TerminalWorkspace workspace,
         VaultKnownHostStore knownHosts,
         SshNetConnectionFactory connections,
-        Func<string, Task> copyToClipboard)
+        Func<string, Task> copyToClipboard,
+        Func<Task<string?>> readClipboard)
     {
         // Difference 3: the Android keystore, with a fingerprint or the device credential releasing the
         // key. A straight implementation of the interface the session layer has always taken.
@@ -207,6 +263,7 @@ public sealed partial class DodoSshApp : Avalonia.Application
             deviceName: PhoneEnvironment.DeviceName,
 
             copyToClipboard: copyToClipboard,
+            readClipboard: readClipboard,
 
             updates: updates);
 

src/DodoSSH.Client.Android/DodoSshAndroidApplication.cs

@@ -1,6 +1,7 @@
 // See the note at the top of MainActivity for why the platform namespaces are reached through `global::`.
 using Avalonia;
 using Avalonia.Android;
+using Avalonia.Media;
 using DodoSSH.Client.Android.Platform;
 using global::Android.App;
 using global::Android.Runtime;
@@ -46,6 +47,30 @@ public sealed class DodoSshAndroidApplication : AvaloniaAndroidApplication<DodoS
     }
 
     /// <inheritdoc />
+    /// <remarks>
+    /// This used to be the one place the two heads disagreed on purpose: the recorded v2 decision was
+    /// "Android recolours with the shared palette and keeps its own default sans", on the reasoning that a
+    /// phone's system font is a phone's own business. The user has reversed that, this pass — the phone now
+    /// takes the desktop's face as well as its colours, and <c>docs/design-import-gaps.md</c> is corrected
+    /// to say so rather than left asserting a decision that no longer holds.
+    ///
+    /// <c>WithInterFont</c> still registers Inter, for the same reason <c>Program.cs</c> keeps it on the
+    /// desktop: it is what the layout suite pins and what a glyph Montserrat does not cover falls back to.
+    /// What changes is which face answers first. Montserrat ships embedded in
+    /// <c>DodoSSH.Client.Shell/Assets/Fonts</c> already — this project references that assembly for
+    /// <c>Theme/Phone.axaml</c>'s own <c>MonoFont</c>, so no csproj or asset work was needed to reach it
+    /// from here, only the same <see cref="FontManagerOptions"/> block the desktop's
+    /// <c>Program.BuildAvaloniaApp</c> sets.
+    /// </remarks>
     protected override AppBuilder CustomizeAppBuilder(AppBuilder builder) =>
-        base.CustomizeAppBuilder(builder).WithInterFont();
+        base.CustomizeAppBuilder(builder)
+            .WithInterFont()
+            .With(new FontManagerOptions
+            {
+                DefaultFamilyName = "avares://DodoSSH.Client.Shell/Assets/Fonts#Montserrat",
+                FontFallbacks =
+                [
+                    new FontFallback { FontFamily = new FontFamily("avares://Avalonia.Fonts.Inter/Assets#Inter") },
+                ],
+            });
 }

src/DodoSSH.Client.Android/MainActivity.cs

@@ -90,6 +90,11 @@ public sealed class MainActivity : AvaloniaMainActivity
         PhoneEnvironment.CurrentActivity = this;
 
         AndroidRedirectCallback.Abandon();
+
+        // And the keyboard may have gone away while this activity was not on screen, without Avalonia's
+        // input pane raising anything to say so. PhoneShell re-reads the inset from this. See
+        // PhoneEnvironment.Resumed, which carries the whole of why.
+        PhoneEnvironment.RaiseResumed();
     }
 
     /// <inheritdoc />

src/DodoSSH.Client.Android/Platform/PhoneEnvironment.cs

@@ -100,6 +100,35 @@ internal static class PhoneEnvironment
     /// </remarks>
     public static global::Android.App.Activity? CurrentActivity { get; set; }
 
+    /// <summary>Raised each time the launcher activity comes back to the foreground.</summary>
+    /// <remarks>
+    /// <para>
+    /// ◆ <b>Exists because the software keyboard can go away without Avalonia being told.</b>
+    /// <c>IInputPane.StateChanged</c> is the only signal <c>PhoneShell</c> has for the keyboard opening and
+    /// closing, and it is not raised when Android dismisses the keyboard on the way out of the app — the
+    /// activity is being paused, and the pane is simply gone by the time anything could have reported it.
+    /// So the inset that holds the interface clear of the keyboard was still applied on the way back in,
+    /// and the whole phone drew an entire keyboard's height above the bottom of a screen with no keyboard
+    /// on it. Reproduced every time: open the keyboard anywhere, press Home, return — 838 of 2340 pixels
+    /// dead, and staying that way until the keyboard was opened and closed again.
+    /// </para>
+    /// <para>
+    /// An event here rather than a call into the view, because the view is Avalonia's and this is Android's:
+    /// <c>MainActivity</c> already owns <c>OnResume</c> and already writes <see cref="CurrentActivity"/>
+    /// from it, and a static event is the same shape of hand-off one step further. Anything else that needs
+    /// to re-read platform state on the way back in can subscribe without <c>MainActivity</c> learning what
+    /// it is.
+    /// </para>
+    /// <para>
+    /// Subscribers are responsible for unsubscribing — this is a static event and holds whatever it is
+    /// given for the life of the process. <c>PhoneShell</c> attaches and detaches with its visual tree.
+    /// </para>
+    /// </remarks>
+    public static event EventHandler? Resumed;
+
+    /// <summary>Announces that the launcher activity is foreground again.</summary>
+    internal static void RaiseResumed() => Resumed?.Invoke(null, EventArgs.Empty);
+
     /// <summary>The application context, once <see cref="Attach"/> has run.</summary>
     public static Context Require() =>
         context ?? throw new InvalidOperationException(

src/DodoSSH.Client.Android/Platform/RendererFocus.cs

@@ -0,0 +1,155 @@
+using global::Android.Views;
+
+namespace DodoSSH.Client.Android.Platform;
+
+/// <summary>
+/// Takes the keyboard away from the terminal's <c>WebView</c>, and reports the taps that should bring it
+/// back.
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>This replaces <c>TerminalFocus</c>, and it is the opposite move.</b> That class handed Android's own
+/// view focus <em>back</em> to the <c>WebView</c> after Avalonia's <c>DispatchTouchEvent</c> took it, because
+/// the input connection the keyboard was attached to belonged to the page. Under ADR 0015 it no longer does:
+/// <see cref="Views.TerminalInputSurface"/> is the input target, and it is an Avalonia control, so the focus
+/// the platform requests for Avalonia's view after every handled touch is now the focus we want. The fight
+/// ends by removing one of the contestants rather than by refereeing it faster.
+/// </para>
+/// <para>
+/// <b>Non-focusable, not non-interactive, and the distinction is the whole design.</b> Focusability governs
+/// where keyboard and D-pad input go; it has nothing to do with touch delivery. A <c>WebView</c> with
+/// <c>Focusable=false</c> still receives, scrolls and hit-tests every touch it is given — so the page keeps
+/// its <c>touch-action: pan-y</c> scrollback exactly as before — but it can no longer take an input
+/// connection, and Chromium's own <c>showSoftInput</c> for the focused DOM node cannot succeed from an
+/// unfocused view. That is precisely the split wanted: the page keeps the finger, Avalonia keeps the
+/// keyboard.
+/// </para>
+/// <para>
+/// <b>Why a tap has to be reported at all.</b> With the page unable to raise a keyboard, the ordinary way
+/// back after the system back gesture lowers one — touch the terminal — would otherwise do nothing. The
+/// listener is attached with <c>Handled = false</c> throughout, so it observes without consuming: the
+/// <c>WebView</c>'s own <c>OnTouchEvent</c> runs after it exactly as if nothing were listening.
+/// </para>
+/// <para>
+/// A tap is distinguished from a scroll the way every Android view does it, by
+/// <c>ViewConfiguration.ScaledTouchSlop</c>. Without that test, dragging the scrollback would raise the
+/// keyboard over the very output being scrolled to.
+/// </para>
+/// <para>
+/// Found by walking the decor view rather than asked of the <c>NativeWebView</c> control, for the reason
+/// <c>TerminalFocus</c> gave and which has not changed: the control does not expose its platform child, and
+/// this application only ever has the one <c>WebView</c>, so the walk's first match is necessarily the
+/// terminal. Every step is allowed to be absent — there is no activity while backgrounded, and no
+/// <c>WebView</c> until the terminal surface has been shown once.
+/// </para>
+/// </remarks>
+internal static class RendererFocus
+{
+    /// <summary>The view already configured, so that showing the terminal twice does not listen twice.</summary>
+    /// <remarks>
+    /// Weak, because this is a static holding an Android view: a strong reference would keep the whole view
+    /// hierarchy, and the activity behind it, alive for the life of the process.
+    /// </remarks>
+    private static WeakReference<global::Android.Webkit.WebView>? configured;
+
+    /// <summary>
+    /// Makes the renderer unable to hold the keyboard, and calls <paramref name="onTapped"/> when somebody
+    /// taps it.
+    /// </summary>
+    /// <remarks>
+    /// Idempotent, and called on every transition into the terminal rather than once at startup, because the
+    /// <c>WebView</c> does not exist until the surface it lives on has been shown — and because Android is
+    /// free to tear its renderer down and build a new one, which is a case this head already handles
+    /// elsewhere (see <c>docs/android-port.md</c>).
+    /// </remarks>
+    public static void Suppress(Action onTapped)
+    {
+        if (PhoneEnvironment.CurrentActivity?.Window?.DecorView is not ViewGroup decor)
+        {
+            return;
+        }
+
+        if (FindWebView(decor) is not { } webView)
+        {
+            return;
+        }
+
+        if (configured is not null && configured.TryGetTarget(out var already) && ReferenceEquals(already, webView))
+        {
+            return;
+        }
+
+        webView.Focusable = false;
+        webView.FocusableInTouchMode = false;
+        webView.ClearFocus();
+
+        WatchForTaps(webView, onTapped);
+
+        configured = new WeakReference<global::Android.Webkit.WebView>(webView);
+    }
+
+    /// <summary>
+    /// Calls <paramref name="onTapped"/> for a tap, and stays out of the way of everything else.
+    /// </summary>
+    /// <remarks>
+    /// A tap is a press and a release that did not travel further than the platform's own
+    /// <c>ScaledTouchSlop</c> — the same test every Android view uses, and necessary rather than fastidious:
+    /// without it, dragging the scrollback would raise the keyboard over the output being scrolled to.
+    /// </remarks>
+    private static void WatchForTaps(global::Android.Webkit.WebView webView, Action onTapped)
+    {
+        // The fallback is the platform's own historical default, for a context that has somehow gone.
+        var slop = webView.Context is { } context ? ViewConfiguration.Get(context)?.ScaledTouchSlop ?? 8 : 8;
+        var downX = 0f;
+        var downY = 0f;
+        var moved = false;
+
+        webView.Touch += (_, e) =>
+        {
+            // Observed, never consumed. The WebView's own OnTouchEvent still runs, so the page scrolls.
+            e.Handled = false;
+
+            if (e.Event is not { } motion)
+            {
+                return;
+            }
+
+            switch (motion.Action)
+            {
+                case MotionEventActions.Down:
+                    downX = motion.GetX();
+                    downY = motion.GetY();
+                    moved = false;
+
+                    break;
+
+                case MotionEventActions.Move:
+                    moved |= Math.Abs(motion.GetX() - downX) > slop || Math.Abs(motion.GetY() - downY) > slop;
+
+                    break;
+
+                case MotionEventActions.Up when !moved:
+                    onTapped();
+
+                    break;
+            }
+        };
+    }
+
+    private static global::Android.Webkit.WebView? FindWebView(ViewGroup parent)
+    {
+        for (var i = 0; i < parent.ChildCount; i++)
+        {
+            switch (parent.GetChildAt(i))
+            {
+                case global::Android.Webkit.WebView webView:
+                    return webView;
+
+                case ViewGroup child when FindWebView(child) is { } found:
+                    return found;
+            }
+        }
+
+        return null;
+    }
+}

src/DodoSSH.Client.Android/Platform/SessionForegroundService.cs

@@ -6,7 +6,7 @@ using global::Android.OS;
 namespace DodoSSH.Client.Android.Platform;
 
 /// <summary>
-/// Keeps the process alive for as long as a shell or a transfer is live.
+/// Keeps the process alive for as long as a shell, a transfer, or a connected Files session is live.
 /// </summary>
 /// <remarks>
 /// <para>
@@ -41,6 +41,26 @@ internal sealed class SessionForegroundService : Service
     private const string ChannelId = "dodossh.sessions";
     private const int NotificationId = 1;
 
+    // API 33+ requires the request to name a code the RequestPermissionsResult callback would be handed
+    // back — 1 is fine because this head never implements that callback at all, see
+    // RequestNotificationPermission's own remark for why a result is not worth listening for.
+    private const int NotificationPermissionRequestCode = 1;
+
+    /// <summary>
+    /// Whether <see cref="OnStartCommand"/> has run for this process without a matching
+    /// <see cref="OnDestroy"/> since — i.e. whether Android currently considers this service foregrounded.
+    /// </summary>
+    /// <remarks>
+    /// Volatile because <see cref="Reconcile"/> can run on whatever thread called
+    /// <see cref="SessionKeepAlive.Refresh"/>, while this is set from the binder thread Android delivers
+    /// service lifecycle callbacks on — two threads with no other synchronisation between them, and a stale
+    /// read here is the difference between updating a notification in place and calling
+    /// <c>StartForegroundService</c> on a service that is already running, which is what defect 3 was.
+    /// </remarks>
+    private static volatile bool running;
+
+    private static bool notificationPermissionRequested;
+
     /// <remarks>
     /// A bound service would tie the sessions' lifetime to a binding, which is the opposite of what is
     /// wanted here: the point is that they outlive whatever the user does with the interface.
@@ -49,7 +69,9 @@ internal sealed class SessionForegroundService : Service
 
     public override StartCommandResult OnStartCommand(Intent? intent, StartCommandFlags flags, int startId)
     {
-        StartForeground(NotificationId, BuildNotification(intent?.GetStringExtra("summary") ?? "Working"));
+        running = true;
+
+        StartForeground(NotificationId, BuildNotification(this, intent?.GetStringExtra("summary") ?? "Working"));
 
         // NotSticky: if Android does kill this process, the SSH connections died with it and there is
         // nothing to resume. Restarting the service would produce a notification claiming sessions that no
@@ -58,14 +80,36 @@ internal sealed class SessionForegroundService : Service
         return StartCommandResult.NotSticky;
     }
 
+    /// <inheritdoc />
     /// <remarks>
+    /// The other half of <see cref="running"/>. Android calls this whether the service stopped itself or
+    /// was stopped from outside — <see cref="Reconcile"/>'s down case calls <c>StopService</c> rather than
+    /// clearing the flag directly, so this override is the one place that actually knows the service has
+    /// gone, matching how <see cref="OnStartCommand"/> is the one place that knows it has come up.
+    /// </remarks>
+    public override void OnDestroy()
+    {
+        running = false;
+
+        base.OnDestroy();
+    }
+
+    /// <remarks>
+    /// <para>
     /// Low importance on purpose. This notification is a receipt, not an alert — it exists because Android
     /// requires one, and because the user is entitled to know the app is holding connections open. Making
     /// it buzz would be a notification about nothing having happened.
+    /// </para>
+    /// <para>
+    /// Static, taking the <see cref="Context"/> it needs rather than reading <c>this</c>: the instance path
+    /// through <see cref="OnStartCommand"/> passes the service itself, and the in-place update path through
+    /// <see cref="Reconcile"/> has no service instance at all — only <see cref="PhoneEnvironment.Require"/>
+    /// — because posting to an already-running notification never touches the service's own lifecycle.
+    /// </para>
     /// </remarks>
-    private Notification BuildNotification(string summary)
+    private static Notification BuildNotification(Context context, string summary)
     {
-        var manager = (NotificationManager)GetSystemService(NotificationService)!;
+        var manager = (NotificationManager)context.GetSystemService(Context.NotificationService)!;
 
         if (OperatingSystem.IsAndroidVersionAtLeast(26))
         {
@@ -79,12 +123,12 @@ internal sealed class SessionForegroundService : Service
         }
 
         var reopen = PendingIntent.GetActivity(
-            this,
+            context,
             0,
-            new Intent(this, typeof(MainActivity)).SetFlags(ActivityFlags.SingleTop),
+            new Intent(context, typeof(MainActivity)).SetFlags(ActivityFlags.SingleTop),
             PendingIntentFlags.Immutable | PendingIntentFlags.UpdateCurrent);
 
-        return new Notification.Builder(this, ChannelId)
+        return new Notification.Builder(context, ChannelId)
             .SetContentTitle("DodoSSH")
             .SetContentText(summary)
             .SetSmallIcon(global::Android.Resource.Drawable.IcDialogInfo)
@@ -93,37 +137,138 @@ internal sealed class SessionForegroundService : Service
             .Build();
     }
 
-    /// <summary>Starts or stops the service to match what is actually running.</summary>
+    /// <summary>Starts, stops, or refreshes the service's notification to match what is actually running.</summary>
     /// <param name="liveSessions">Shells with a live channel behind them.</param>
     /// <param name="activeTransfers">Transfers still moving bytes.</param>
-    public static void Reconcile(int liveSessions, int activeTransfers)
+    /// <param name="holdsFileSession">Whether the Files screen holds a live, idle SFTP connection.</param>
+    /// <remarks>
+    /// Three outcomes, not the two a plain start-or-stop would have. Nothing live stops the service, as
+    /// always. Something live and the service not yet running starts it. Something live and the service
+    /// already running is the case that used to call <c>StartForegroundService</c> a second time, which on
+    /// API 31+ throws <c>ForegroundServiceStartNotAllowedException</c> the instant the app is backgrounded
+    /// — a transfer finishing in the pocket, one of two shells dying — crashing the process and taking the
+    /// remaining connections with it. That case — running, and the app backgrounded — now only posts a
+    /// fresh notification through the <see cref="NotificationManager"/> already holding the channel open,
+    /// which needs no foreground-start permission at all. A foregrounded refresh still prefers a real
+    /// start even over a service that looks like it is running; the inline remark below is why.
+    /// </remarks>
+    public static void Reconcile(int liveSessions, int activeTransfers, bool holdsFileSession)
     {
         var context = PhoneEnvironment.Require();
-        var intent = new Intent(context, typeof(SessionForegroundService));
 
-        if (liveSessions == 0 && activeTransfers == 0)
+        if (liveSessions == 0 && activeTransfers == 0 && !holdsFileSession)
         {
-            context.StopService(intent);
+            context.StopService(new Intent(context, typeof(SessionForegroundService)));
 
             return;
         }
 
         // The summary says what is actually held, counted rather than generic — the same principle the
         // delete confirmations follow. "DodoSSH is running" would tell the user nothing they could act on.
-        intent.PutExtra("summary", Summarise(liveSessions, activeTransfers));
+        var summary = Summarise(liveSessions, activeTransfers, holdsFileSession);
 
-        context.StartForegroundService(intent);
+        // In-place only while backgrounded, where it is the only legal move. Foregrounded, a real start is
+        // always allowed and is preferred even when `running` says the service is up: a disconnect followed
+        // by a quick reconnect can land here while the StopService just issued is still in flight, and
+        // posting to that dying service's notification would leave an orphan receipt over an unprotected
+        // process — restarting instead makes the flag's small lag harmless. A start on a service that
+        // really is running only re-delivers OnStartCommand, whose StartForeground updates the same
+        // notification anyway. CurrentActivity is the foreground signal: set on resume, cleared on pause.
+        if (running && PhoneEnvironment.CurrentActivity is null)
+        {
+            var manager = (NotificationManager)context.GetSystemService(Context.NotificationService)!;
+            manager.Notify(NotificationId, BuildNotification(context, summary));
+
+            return;
+        }
+
+        RequestNotificationPermission(context);
+        Start(context, summary);
     }
 
-    private static string Summarise(int liveSessions, int activeTransfers)
+    /// <remarks>
+    /// Split out of <see cref="Reconcile"/> for the try/catch alone, which needs its own remark and would
+    /// otherwise crowd the three-way branch above it.
+    /// </remarks>
+    private static void Start(Context context, string summary)
     {
-        var parts = new List<string>(2);
+        var intent = new Intent(context, typeof(SessionForegroundService));
+        intent.PutExtra("summary", summary);
+
+        try
+        {
+            context.StartForegroundService(intent);
+        }
+        catch (Java.Lang.IllegalStateException)
+        {
+            // ForegroundServiceStartNotAllowedException (API 31+) derives from this, and reaching it here
+            // means a start-worthy transition — a shell opening, a Files connection completing, the first
+            // transfer landing in an empty queue — happened while the app was backgrounded, which is
+            // precisely when Android refuses a new foreground start. There is no retry that helps: by the
+            // time this catch runs, the moment such a start would have been allowed has already passed.
+            // Swallowing it is the honest choice and not just the available one — letting the exception
+            // propagate would crash the process and drop the very shells and transfers this service exists
+            // to keep alive. A process that keeps running unprotected outlives one that does not run at all.
+        }
+    }
+
+    /// <summary>
+    /// Asks for the receipt notification's own permission, the first time this process actually has
+    /// something to show rather than at launch.
+    /// </summary>
+    /// <remarks>
+    /// At most once per process, via <see cref="notificationPermissionRequested"/> — not to work around a
+    /// platform limit, since Android already refuses to show the dialogue twice, but because a second call
+    /// to <c>RequestPermissions</c> after the first is still pending is its own kind of noise. No result is
+    /// read back: there is nothing this class would do differently for a grant versus a refusal, so a
+    /// callback would exist only to be empty. What refusal costs is stated rather than hidden — the
+    /// notification stays invisible — and what it does not cost is the point: the service still starts,
+    /// still holds the process in the foreground, and the shells and transfers it protects are exactly as
+    /// safe as if the user had said yes. See the manifest's own comment on this permission.
+    /// </remarks>
+    private static void RequestNotificationPermission(Context context)
+    {
+        // Isolated as its own guard clause rather than folded into the compound condition below: the
+        // platform-compatibility analyzer only recognises a version check as guarding what follows when it
+        // is the sole condition of its own early return, and PostNotifications is annotated API 33+.
+        if (!OperatingSystem.IsAndroidVersionAtLeast(33))
+        {
+            return;
+        }
+
+        // Read into a local rather than referenced again inside the lambda below: the guard clause above
+        // covers a direct call in this method's own body, but the platform-compatibility analyzer treats a
+        // lambda as reachable from anywhere and will not extend the guard across that boundary. Capturing
+        // the already-validated string sidesteps the false positive without weakening the actual check.
+        var postNotifications = global::Android.Manifest.Permission.PostNotifications;
+
+        if (notificationPermissionRequested
+            || PhoneEnvironment.CurrentActivity is not { } activity
+            || context.CheckSelfPermission(postNotifications) == Permission.Granted)
+        {
+            return;
+        }
+
+        notificationPermissionRequested = true;
+
+        activity.RunOnUiThread(() =>
+            activity.RequestPermissions([postNotifications], NotificationPermissionRequestCode));
+    }
+
+    private static string Summarise(int liveSessions, int activeTransfers, bool holdsFileSession)
+    {
+        var parts = new List<string>(3);
 
         if (liveSessions > 0)
         {
             parts.Add(liveSessions == 1 ? "1 shell connected" : $"{liveSessions} shells connected");
         }
 
+        if (holdsFileSession)
+        {
+            parts.Add("Files connected");
+        }
+
         if (activeTransfers > 0)
         {
             parts.Add(activeTransfers == 1 ? "1 transfer running" : $"{activeTransfers} transfers running");

src/DodoSSH.Client.Android/Platform/SessionKeepAlive.cs

@@ -17,37 +17,53 @@ namespace DodoSSH.Client.Android.Platform;
 /// tally kept here. It already knows that a session whose shell exited half an hour ago is not live, which
 /// a counter incremented on open and decremented on close would not.
 /// </para>
+/// <para>
+/// Three facts feed <see cref="SessionForegroundService.Reconcile"/>, not one: live shells, moving
+/// transfers, and an idle-but-connected Files session. The last of those used to be missing entirely —
+/// a shell survives backgrounding because somebody opened it, but a Files connection with nothing moving
+/// looked, to this class, exactly like nothing being open at all. <c>holdsFileSession</c> below is that
+/// gap closed, read the same way the other two facts are: asked, not cached.
+/// </para>
 /// </remarks>
 internal sealed class SessionKeepAlive : IDisposable
 {
     private readonly TerminalWorkspace workspace;
     private readonly Func<int> activeTransfers;
+    private readonly Func<bool> holdsFileSession;
 
     /// <param name="workspace">The live shells.</param>
     /// <param name="activeTransfers">
-    /// How many transfers are moving bytes. A delegate rather than a queue, because file transfer is out
-    /// of this head's first scope — see the decision in docs/android-port.md — and this is the seam it
-    /// will arrive through rather than a dependency taken before there is anything to depend on.
+    /// How many transfers are moving bytes. A delegate rather than a queue, because ownership of the
+    /// transfer queue stays with <c>TransfersViewModel</c> — this class only ever asks it a question.
     /// </param>
-    public SessionKeepAlive(TerminalWorkspace workspace, Func<int> activeTransfers)
+    /// <param name="holdsFileSession">
+    /// Whether the Files screen holds a live SFTP connection with nothing moving on it — the idle-but-
+    /// connected case a transfer count alone would miss. See <c>TransfersViewModel.HasLiveFileSession</c>.
+    /// </param>
+    public SessionKeepAlive(TerminalWorkspace workspace, Func<int> activeTransfers, Func<bool> holdsFileSession)
     {
         this.workspace = workspace;
         this.activeTransfers = activeTransfers;
+        this.holdsFileSession = holdsFileSession;
 
-        // Raised on whatever thread the pump unwound on, which is fine: starting and stopping a service is
-        // a binder call and needs no particular thread. Nothing here touches the interface.
+        // Raised on whatever thread the workspace announced from — a continuation of the ended run, or the
+        // closer's own — which is fine: starting and stopping a service is a binder call and needs no
+        // particular thread. Nothing here touches the interface. That the announcement waits for the run to
+        // actually complete, and comes for deliberate closes too, is what makes reading LiveSessionCount
+        // from it honest — the event's own remark carries the stuck notification that taught us both.
         workspace.SessionEnded += OnSessionEnded;
     }
 
     /// <summary>Re-reads the counts and starts or stops the service to match.</summary>
     /// <remarks>
-    /// Called after anything that could change either count — opening a shell, closing a tab, a transfer
-    /// finishing. Calling it when nothing changed is free: reconciling to the state it is already in is
-    /// either a redundant <c>startForegroundService</c> on a running service or a <c>stopService</c> on a
-    /// stopped one, and Android treats both as no-ops.
+    /// Called after anything that could change any of the three facts — opening a shell, closing a tab, a
+    /// transfer finishing, a Files connection opening or closing. Calling it when nothing changed is free:
+    /// reconciling to the state it is already in is either a redundant <c>startForegroundService</c> — or,
+    /// now, a redundant notification post — on a running service, or a <c>stopService</c> on a stopped one,
+    /// and Android treats all of those as no-ops.
     /// </remarks>
     public void Refresh() =>
-        SessionForegroundService.Reconcile(workspace.LiveSessionCount, activeTransfers());
+        SessionForegroundService.Reconcile(workspace.LiveSessionCount, activeTransfers(), holdsFileSession());
 
     /// <inheritdoc />
     public void Dispose()
@@ -56,7 +72,7 @@ internal sealed class SessionKeepAlive : IDisposable
 
         // The notification goes with the composition root. Leaving it up over a process that is shutting
         // down is how an SSH client acquires a reputation for a notification you cannot get rid of.
-        SessionForegroundService.Reconcile(0, 0);
+        SessionForegroundService.Reconcile(0, 0, false);
     }
 
     private void OnSessionEnded(object? sender, TerminalSessionEndedEventArgs e) => Refresh();

src/DodoSSH.Client.Android/Properties/AndroidManifest.xml

@@ -12,7 +12,12 @@
   <uses-permission android:name="android.permission.FOREGROUND_SERVICE" />
   <uses-permission android:name="android.permission.FOREGROUND_SERVICE_DATA_SYNC" />
 
-  <!-- The service's persistent notification. Runtime-requested on API 33+, and refusal is survivable. -->
+  <!--
+    The service's persistent notification. Requested on API 33+ from SessionForegroundService.Reconcile,
+    the first time in this process there is actually something to show — not at launch, where the ask
+    would justify nothing on screen yet. Refusal is survivable: the service still starts and still holds
+    the process in the foreground either way, so a "no" costs the notification and nothing else.
+  -->
   <uses-permission android:name="android.permission.POST_NOTIFICATIONS" />
 
   <!-- Releases the device key. See AndroidDeviceKeyStore. -->

src/DodoSSH.Client.Android/Resources/drawable/ic_launcher_foreground.xml

@@ -1,7 +1,7 @@
 <?xml version="1.0" encoding="utf-8"?>
 <!--
   The launcher mark, and it is the same mark PhoneShell's header and the desktop titlebar draw:
-  >_ in the canvas colour on a solid accent tile. Filled rather than outlined since v2.
+  >_ in AccentInk on a solid accent tile. Filled rather than outlined since v2.
 
   The tile is not in this file. It is the background layer — @color/dodo_accent, see ic_launcher.xml
   — and that is the whole trick of the filled design on Android: the rounding a launcher applies is
@@ -10,10 +10,10 @@
   rounded shape inside the first, visibly clipped at the corners on any device whose mask is not the
   one it was drawn for.
 
-  #0E1220 is AccentInk from DodoSSH.Client.Shell's Theme/Palette.axaml, written out because an
-  Android resource cannot reference a XAML dictionary. It equals Canvas today and is named
-  separately in the palette for a reason worth keeping in mind here: this is ink on the accent, not
-  the window behind it, and it follows AccentInk if the two ever part.
+  #FFFFFF is AccentInk from DodoSSH.Client.Shell's Theme/Palette.axaml, written out because an
+  Android resource cannot reference a XAML dictionary. Through v2 it equalled Canvas and this file
+  followed it when the two parted: v5's ink on the accent is white, the way every accent surface in
+  that palette carries white, while the window behind it went its own way to #05050A.
 
   108x108 with the artwork inside the middle 72 is the adaptive-icon contract — the outer 18 on each
   edge is what the launcher eats for masking and parallax. The glyph spans 36.06..71.94, which is
@@ -35,7 +35,7 @@
   <path
       android:pathData="M36.06,43.33 L49.91,53.57 L36.06,63.8"
       android:fillColor="#00000000"
-      android:strokeColor="#0E1220"
+      android:strokeColor="#FFFFFF"
       android:strokeWidth="5.4"
       android:strokeLineCap="round"
       android:strokeLineJoin="round" />
@@ -44,7 +44,7 @@
   <path
       android:pathData="M54,64.68 L71.94,64.68"
       android:fillColor="#00000000"
-      android:strokeColor="#0E1220"
+      android:strokeColor="#FFFFFF"
       android:strokeWidth="5.4"
       android:strokeLineCap="round" />
 

src/DodoSSH.Client.Android/Resources/drawable/ic_launcher_monochrome.xml

@@ -11,7 +11,7 @@
 
   Worth shipping rather than leaving out: a launcher with themed icons on and no monochrome
   layer to use falls back to the full-colour icon, so the one app on the home screen still
-  drawn in blue is this one.
+  drawn in its own accent is this one.
 -->
 <vector xmlns:android="http://schemas.android.com/apk/res/android"
     android:width="108dp"

src/DodoSSH.Client.Android/Resources/values/colors.xml

@@ -9,12 +9,12 @@
     between an Android resource and a XAML resource dictionary, so the duplication is stated rather than
     hidden.
   -->
-  <color name="dodo_window">#0E1220</color>
+  <color name="dodo_window">#05050A</color>
 
   <!--
     AccentColor from the same palette, here because the launcher icon's background layer is a colour
     and not a drawable. Same hand-kept duplication as above, and the same rule: if the palette moves,
     this moves with it.
   -->
-  <color name="dodo_accent">#5B8CFF</color>
+  <color name="dodo_accent">#5D42DE</color>
 </resources>

src/DodoSSH.Client.Android/Theme/Phone.axaml

@@ -12,10 +12,28 @@
     the alternative, and the red one is the one that costs something.
 
     ── v2 ──────────────────────────────────────────────────────────────────────────────────────────────
-    The second design rounds everything. The corner radii below are the design's own — 4 for a tag, 9 for a
-    button or a pill, 10 for a list row, 11 for the search well, 12 for a card, 14 for a block of
-    monospaced output — and they are a ladder rather than a set of preferences: the radius says how big the
-    thing is, so a 12 on a chip or a 4 on a card reads as the wrong size before it reads as the wrong shape.
+    The second design rounded everything on its own ladder — 4 for a tag, 9 for a button or a pill, 10 for
+    a list row, 11 for the search well, 12 for a card, 14 for a block of monospaced output — and the numbers
+    below carried it for three passes: the radius said how big the thing was, so a 12 on a chip or a 4 on a
+    card read as the wrong size before it read as the wrong shape. Left as a record of that reasoning rather
+    than deleted, because the reasoning still holds; only the numbers it was reasoning about have moved.
+
+    ── v5 ──────────────────────────────────────────────────────────────────────────────────────────────
+    This pass takes the desktop's v5 ladder instead of v2's own, on the same reversal recorded for the
+    fonts: the phone now matches the desktop's shape as well as its face. Three rungs rather than six —
+    cards and sections stay 12, a button or a field is 10, a chip or a tag is 6 — collapsing v2's 9/10/11
+    into the one value the desktop's buttons and fields already use, and moving its 4 up to 6 and its 14
+    down to 12 to land on the desktop's own chip and card numbers. The ladder still says how big a thing
+    is before it says what shape it is; it is just a shorter ladder now, because the desktop it is copying
+    never drew v2's 11-radius search well or a card any rounder than a chip's neighbour a step away, and the
+    same case that closed the gap between 9, 10 and 11 closes the one between 12 and 14.
+
+    Not every radius on this head belongs to this ladder. The floating action button is 28 — half its own
+    56, which is a circle rather than a ladder rung — and the sheets stay at 22 on their top corners only,
+    which is a phone idiom this codebase's own bottom sheets have used since v2 and the desktop draws
+    nothing like. Both are documented where they are set rather than here, for the reason FAB and sheet
+    radii are always documented locally: a reader who only ever meets one of them should not have to find
+    this paragraph to learn it was deliberate.
   -->
 
   <Style Selector="Button.primary">
@@ -23,15 +41,32 @@
     <Setter Property="HorizontalAlignment" Value="Stretch" />
     <Setter Property="HorizontalContentAlignment" Value="Center" />
     <Setter Property="VerticalContentAlignment" Value="Center" />
-    <Setter Property="Background" Value="{StaticResource Accent}" />
     <Setter Property="Foreground" Value="{StaticResource AccentInk}" />
-    <Setter Property="CornerRadius" Value="9" />
+    <Setter Property="CornerRadius" Value="10" />
     <Setter Property="FontFamily" Value="{StaticResource MonoFont}" />
     <Setter Property="FontSize" Value="12" />
     <Setter Property="FontWeight" Value="SemiBold" />
   </Style>
+
+  <!--
+    ◆ Gradient and glow, replacing a flat Accent fill — the same swap App.axaml's Button.accent made on the
+    desktop, and for the same reason: this design draws every primary button top-to-bottom from a lighter
+    violet into the accent proper, lifted off the surface with a soft violet glow rather than a border. Both
+    live in Palette.axaml as AccentGradient and AccentGlow, already resolvable here because the palette is
+    shared — this is a resource-key swap, not new colour.
+
+    Background and BoxShadow are set here rather than as plain Setters above, because Fluent's default
+    button template only lets a style reach the fill and the glow through the ContentPresenter it draws
+    itself around — the same reason the desktop's rule targets /template/ ContentPresenter rather than the
+    Button. Height, radius and the rest stay ordinary Setters on Button.primary itself; only the two the
+    template intercepts move down here.
+  -->
+  <Style Selector="Button.primary /template/ ContentPresenter">
+    <Setter Property="Background" Value="{StaticResource AccentGradient}" />
+    <Setter Property="BoxShadow" Value="{StaticResource AccentGlow}" />
+  </Style>
   <Style Selector="Button.primary:pressed /template/ ContentPresenter">
-    <Setter Property="Background" Value="{StaticResource Accent}" />
+    <Setter Property="Background" Value="{StaticResource AccentGradient}" />
     <Setter Property="Opacity" Value="0.82" />
   </Style>
 
@@ -39,9 +74,15 @@
     Disabled is drawn as flat and unlit rather than merely dimmed. The design's CONTINUE button on the
     recovery screen is disabled until the checkbox is ticked, and a user who cannot tell it is disabled
     reads the screen as broken rather than as waiting for them.
+
+    BoxShadow has to be cleared here too, as "none" rather than left unset — see the remark on
+    Button.accent:disabled in the desktop's App.axaml for why the literal string is required and an empty
+    BoxShadows is not: the base rule's glow Setter is still in effect wherever a more specific one does not
+    override it, and a disabled primary button lit with a glow would read as wanting to be pressed.
   -->
   <Style Selector="Button.primary:disabled /template/ ContentPresenter">
     <Setter Property="Background" Value="{StaticResource Raised}" />
+    <Setter Property="BoxShadow" Value="none" />
     <Setter Property="TextElement.Foreground" Value="{StaticResource TextFaint}" />
   </Style>
 
@@ -54,7 +95,7 @@
     <Setter Property="BorderBrush" Value="{StaticResource BorderMid}" />
     <Setter Property="BorderThickness" Value="1" />
     <Setter Property="Foreground" Value="{StaticResource Text}" />
-    <Setter Property="CornerRadius" Value="9" />
+    <Setter Property="CornerRadius" Value="10" />
     <Setter Property="FontFamily" Value="{StaticResource MonoFont}" />
     <Setter Property="FontSize" Value="11.5" />
     <Setter Property="FontWeight" Value="SemiBold" />
@@ -69,17 +110,33 @@
     <Setter Property="BorderBrush" Value="{StaticResource DangerSoft}" />
     <Setter Property="BorderThickness" Value="1" />
     <Setter Property="Foreground" Value="{StaticResource Danger}" />
-    <Setter Property="CornerRadius" Value="9" />
+    <Setter Property="CornerRadius" Value="10" />
     <Setter Property="FontFamily" Value="{StaticResource MonoFont}" />
     <Setter Property="FontSize" Value="10.5" />
     <Setter Property="FontWeight" Value="SemiBold" />
   </Style>
 
-  <!-- A row in a list: the whole row is the target, and it is 54 tall because a thumb is not a mouse. -->
+  <!--
+    A row in a list: the whole row is the target, and it is 54 tall because a thumb is not a mouse.
+
+    ◆ VerticalContentAlignment, because that height is the whole point of this class and Avalonia's default
+    for content alignment is Stretch — so the content presenter stretched the caption to the full row and a
+    TextBlock draws its line at the TOP of what it is given. Most rows here never showed it, having a
+    StackPanel or a Grid of already-centred children in them, which is what made the four that did look
+    like four unrelated mistakes: the breadcrumb chips and the up-one-directory button on FilesScreen, and
+    TerminalScreen's CLOSE THIS TAB, each a bare TextBlock in a row 36 or 44 tall with no vertical padding.
+    Measured at those numbers, the caption sat flush against the top edge with 21 to 33 pixels below it.
+
+    The desktop head's App.axaml carries the same setter on its own shapes for the same reason, and excludes
+    two of them — see the remark on Button.ghost there. Nothing is excluded here: no row's content depends
+    on being stretched, there being no full-height strip inside any of the thirty-three, and the Grids that
+    stop filling hold only children that already centre themselves, so they land where they always did.
+  -->
   <Style Selector="Button.row">
     <Setter Property="MinHeight" Value="54" />
     <Setter Property="HorizontalAlignment" Value="Stretch" />
     <Setter Property="HorizontalContentAlignment" Value="Stretch" />
+    <Setter Property="VerticalContentAlignment" Value="Center" />
     <Setter Property="Background" Value="Transparent" />
     <Setter Property="BorderThickness" Value="0" />
     <Setter Property="CornerRadius" Value="10" />
@@ -110,10 +167,124 @@
     <Setter Property="Foreground" Value="{StaticResource AccentText}" />
     <Setter Property="FontSize" Value="20" />
   </Style>
+  <!--
+    ◆ THE FILL BELOW IS NOT DECORATION. IT IS UNDOING ONE FLUENT PUTS THERE.
+
+    `accent` is not a name this theme invented — Fluent ships a `Button.accent` style of its own, and a
+    button classed `icon accent` matches it. So these three buttons were being painted with Fluent's
+    system-accent slab, #6275AC, a blue that appears nowhere in Palette.axaml and reads as a control from
+    a different application sitting in this one's header. Measured off a screenshot rather than guessed:
+    the fill was #6275AC where the accent is #5D42DE.
+
+    The rule above could not stop it, because it sets Foreground and FontSize and Fluent's fill is on the
+    ContentPresenter — which is also why `Button.icon`'s own Background="Transparent" does not reach it.
+    App.axaml has carried the desktop's answer to exactly this since v5b: override the ContentPresenter,
+    not the Button. This head simply never had one, so the desktop looked right and the phone did not.
+
+    Transparent rather than the desktop's gradient, and the difference is the control rather than an
+    inconsistency. The desktop's Button.accent is a filled primary slab — this head has one of those too,
+    and it is Button.primary. `icon accent` is a 44-pixel glyph in a header row beside other glyphs; what
+    "accent" asks for there is the accent *colour on the mark*, which is what the rule above gives it. A
+    slab would make one header icon a button and its neighbours not.
+
+    Pressed is deliberately not restated: Button.icon:pressed below sets ChromeHover on every icon button
+    and, coming later in this file, wins over this rule for the moment a finger is down. That is the
+    feedback the plain icon buttons give and there is no reason this one should differ.
+  -->
+  <Style Selector="Button.icon.accent /template/ ContentPresenter#PART_ContentPresenter">
+    <Setter Property="Background" Value="Transparent" />
+    <Setter Property="BorderThickness" Value="0" />
+    <Setter Property="Foreground" Value="{StaticResource AccentText}" />
+  </Style>
+  <Style Selector="Button.icon.accent:pointerover /template/ ContentPresenter#PART_ContentPresenter">
+    <Setter Property="Background" Value="Transparent" />
+    <Setter Property="Foreground" Value="{StaticResource AccentText}" />
+  </Style>
+  <Style Selector="Button.icon.accent:disabled /template/ ContentPresenter#PART_ContentPresenter">
+    <Setter Property="Background" Value="Transparent" />
+    <Setter Property="Foreground" Value="{StaticResource TextFaint}" />
+  </Style>
   <Style Selector="Button.icon:pressed /template/ ContentPresenter">
     <Setter Property="Background" Value="{StaticResource ChromeHover}" />
   </Style>
 
+  <!--
+    ============ one key of the terminal's accessory keypad ============
+
+    A class rather than eleven sets of attributes assigned in code, and the move out of the code-behind is
+    most of what makes the latched state honest: Ctrl on and Ctrl off used to be two Background values
+    written by hand in TerminalScreen.axaml.cs, which meant the row named two of this palette's brushes
+    itself and any third state would have named a third. It is `latched` here, and the code toggles a class.
+
+    ── the numbers, and why they are these ─────────────────────────────────────────────────────────────────
+    40 tall, where the row this replaced was 30. Android asks for 48 on a touch target and every terminal on
+    the platform lands near it; 30 was chosen to save a strip of the terminal above and took the keypad past
+    the point of working, which is the wrong end of that trade. 40 with a 2-pixel margin gives 44 of pitch,
+    which is the same figure TextBox.field uses for a box a thumb has to hit.
+
+    HorizontalAlignment Stretch, because these live in UniformGrid cells and the cell is the share of the
+    width the key is entitled to. Left to Center a key would shrink to its label and "|" would be a sliver
+    between two wide neighbours — which is the layout the flexed row exists to stop.
+
+    ── pressed has to be visible, and it was not ───────────────────────────────────────────────────────────
+    These keys are Focusable="False" so that using one does not take the keyboard away from the terminal
+    (see TerminalScreen.axaml.cs), and a button that never takes focus also never shows the focus adorner
+    Fluent would otherwise give it. That left a press with no feedback at all on a control whose entire job
+    is to be pressed while somebody is looking at the line of output above it. The pressed fill below is the
+    only thing that says a key was hit — and on a keypad over a remote shell, "did that send?" is a question
+    worth answering in the interface rather than in the scrollback.
+  -->
+  <Style Selector="Button.hotkey">
+    <Setter Property="Height" Value="40" />
+    <Setter Property="MinHeight" Value="40" />
+    <Setter Property="MinWidth" Value="0" />
+    <Setter Property="Margin" Value="2,0" />
+    <Setter Property="Padding" Value="0" />
+    <Setter Property="CornerRadius" Value="10" />
+    <Setter Property="Background" Value="{StaticResource Panel}" />
+    <Setter Property="BorderBrush" Value="{StaticResource BorderMid}" />
+    <Setter Property="BorderThickness" Value="1" />
+    <Setter Property="Foreground" Value="{StaticResource TextDim}" />
+    <Setter Property="FontFamily" Value="{StaticResource MonoFont}" />
+    <Setter Property="FontSize" Value="12" />
+    <Setter Property="HorizontalAlignment" Value="Stretch" />
+    <Setter Property="VerticalAlignment" Value="Center" />
+    <Setter Property="HorizontalContentAlignment" Value="Center" />
+    <Setter Property="VerticalContentAlignment" Value="Center" />
+  </Style>
+  <Style Selector="Button.hotkey /template/ ContentPresenter#PART_ContentPresenter">
+    <Setter Property="Background" Value="{StaticResource Panel}" />
+    <Setter Property="BorderBrush" Value="{StaticResource BorderMid}" />
+    <Setter Property="BorderThickness" Value="1" />
+    <Setter Property="CornerRadius" Value="10" />
+  </Style>
+  <Style Selector="Button.hotkey:pointerover /template/ ContentPresenter#PART_ContentPresenter">
+    <Setter Property="Background" Value="{StaticResource Panel}" />
+    <Setter Property="BorderBrush" Value="{StaticResource BorderMid}" />
+  </Style>
+  <Style Selector="Button.hotkey:pressed /template/ ContentPresenter#PART_ContentPresenter">
+    <Setter Property="Background" Value="{StaticResource Active}" />
+    <Setter Property="BorderBrush" Value="{StaticResource Accent}" />
+  </Style>
+  <!--
+    Latched Ctrl. The same fill a press gives, held rather than flashed — a modifier that is on and does not
+    look on is how somebody sends ^L to a database prompt believing they typed an l.
+  -->
+  <Style Selector="Button.hotkey.latched">
+    <Setter Property="Foreground" Value="{StaticResource AccentText}" />
+  </Style>
+  <Style Selector="Button.hotkey.latched /template/ ContentPresenter#PART_ContentPresenter">
+    <Setter Property="Background" Value="{StaticResource Active}" />
+    <Setter Property="BorderBrush" Value="{StaticResource Accent}" />
+  </Style>
+  <Style Selector="Button.hotkey:disabled /template/ ContentPresenter#PART_ContentPresenter">
+    <Setter Property="Background" Value="{StaticResource Panel}" />
+    <Setter Property="BorderBrush" Value="{StaticResource Border}" />
+  </Style>
+  <Style Selector="Button.hotkey:disabled">
+    <Setter Property="Foreground" Value="{StaticResource TextFaint}" />
+  </Style>
+
   <!--
     The floating action button the design has asked for since v2, and which this theme refused twice — on
     the grounds that it would open an editor this head had not got, and that an unused style is a claim the
@@ -126,7 +297,17 @@
 
     Accent-filled, which it shares with Button.primary and with nothing else — and it means the same thing
     in both places: the one action on the surface that is not a choice between peers. Circular by radius
-    rather than by a Path, so the pressed state the template draws is the same shape as the button.
+    rather than by a Path, so the pressed state the template draws is the same shape as the button — 28,
+    exactly half its own 56, which is a geometric constraint rather than a ladder rung and does not move
+    with the rest of this file's radii.
+
+    ◆ Gradient and glow since v5, matching Button.primary rather than staying a flat fill once that one
+    moved. The two are the only accent-filled controls on this head and are read as one idea — "the thing
+    this surface wants you to do" — so a flat FAB beside a gradient primary button would be the seam this
+    codebase's palette file keeps warning about, one screenshot over from the button it echoes. The glow
+    reads as well on a floating circle as it does on a bar-anchored rectangle: if anything a control that
+    already floats over content earns a lift more than one sitting in a row of chrome does, so it is kept
+    rather than dropped for the FAB's own shape.
 
     Still only on HOSTS. The design puts a second one on S3 and that editor genuinely does not exist yet, so
     the style being here is not permission to draw one there.
@@ -135,7 +316,6 @@
     <Setter Property="Width" Value="56" />
     <Setter Property="Height" Value="56" />
     <Setter Property="Padding" Value="0" />
-    <Setter Property="Background" Value="{StaticResource Accent}" />
     <Setter Property="BorderThickness" Value="0" />
     <Setter Property="CornerRadius" Value="28" />
     <Setter Property="HorizontalContentAlignment" Value="Center" />
@@ -145,8 +325,14 @@
     <Setter Property="FontSize" Value="24" />
     <Setter Property="FontWeight" Value="SemiBold" />
   </Style>
+  <!-- See the remark above Button.primary's own /template/ ContentPresenter rule: BoxShadow has no home on
+       a Button itself, so the fill moves down here alongside it rather than staying a plain Setter. -->
+  <Style Selector="Button.fab /template/ ContentPresenter">
+    <Setter Property="Background" Value="{StaticResource AccentGradient}" />
+    <Setter Property="BoxShadow" Value="{StaticResource AccentGlow}" />
+  </Style>
   <Style Selector="Button.fab:pressed /template/ ContentPresenter">
-    <Setter Property="Background" Value="{StaticResource Accent}" />
+    <Setter Property="Background" Value="{StaticResource AccentGradient}" />
     <Setter Property="Opacity" Value="0.82" />
   </Style>
 
@@ -161,7 +347,7 @@
     interface glitching rather than as a tap being received.
   -->
   <Style Selector="Button.scrim">
-    <Setter Property="Background" Value="#9E0E1220" />
+    <Setter Property="Background" Value="#9E05050A" />
     <Setter Property="BorderThickness" Value="0" />
     <Setter Property="CornerRadius" Value="0" />
     <Setter Property="Padding" Value="0" />
@@ -169,10 +355,10 @@
     <Setter Property="VerticalAlignment" Value="Stretch" />
   </Style>
   <Style Selector="Button.scrim:pointerover /template/ ContentPresenter">
-    <Setter Property="Background" Value="#9E0E1220" />
+    <Setter Property="Background" Value="#9E05050A" />
   </Style>
   <Style Selector="Button.scrim:pressed /template/ ContentPresenter">
-    <Setter Property="Background" Value="#9E0E1220" />
+    <Setter Property="Background" Value="#9E05050A" />
   </Style>
 
   <!--
@@ -194,7 +380,7 @@
   -->
   <Style Selector="Border.output">
     <Setter Property="Background" Value="{StaticResource TerminalSurface}" />
-    <Setter Property="CornerRadius" Value="14" />
+    <Setter Property="CornerRadius" Value="12" />
     <Setter Property="Padding" Value="12" />
   </Style>
 
@@ -205,7 +391,7 @@
   -->
   <Style Selector="Border.tag">
     <Setter Property="Background" Value="{StaticResource Chip}" />
-    <Setter Property="CornerRadius" Value="4" />
+    <Setter Property="CornerRadius" Value="6" />
     <Setter Property="Padding" Value="7,2" />
     <Setter Property="VerticalAlignment" Value="Center" />
   </Style>
@@ -231,7 +417,7 @@
     <Setter Property="Background" Value="Transparent" />
     <Setter Property="BorderBrush" Value="{StaticResource BorderMid}" />
     <Setter Property="BorderThickness" Value="1" />
-    <Setter Property="CornerRadius" Value="4" />
+    <Setter Property="CornerRadius" Value="6" />
     <Setter Property="Padding" Value="11,0" />
     <Setter Property="Foreground" Value="{StaticResource TextDim}" />
     <Setter Property="VerticalContentAlignment" Value="Center" />
@@ -267,11 +453,17 @@
     Checked is a filled surface with accent *text*, not an accent fill. v2 makes that distinction
     everywhere — see the remark on AccentText in Palette.axaml — and a chip is where it matters most: a row
     of four solid blue lozenges is a row of four things that all look like the primary action.
+
+    ◆ Its own radius rather than Border.tag's or Button.chiptoggle's, despite the name. v2 drew this control
+    on the ladder's button-or-pill rung rather than its tag rung — 9, not 4 — and the v5 pass carries the
+    same rung forward to 10 rather than to Border.tag's 6: it is a much larger control at 34 tall against a
+    tag's line-height, and a radius picked for a 34-pixel chip is not the one a 20-pixel tag needs, whatever
+    the class is called. See the remark on Phone.axaml's own ladder, above.
   -->
   <Style Selector="RadioButton.chip">
     <Setter Property="MinHeight" Value="34" />
     <Setter Property="Padding" Value="13,6" />
-    <Setter Property="CornerRadius" Value="9" />
+    <Setter Property="CornerRadius" Value="10" />
     <Setter Property="Background" Value="Transparent" />
     <Setter Property="BorderBrush" Value="{StaticResource BorderMid}" />
     <Setter Property="BorderThickness" Value="1" />
@@ -314,7 +506,59 @@
     <Setter Property="Background" Value="{StaticResource Field}" />
     <Setter Property="BorderBrush" Value="{StaticResource BorderMid}" />
     <Setter Property="BorderThickness" Value="1" />
-    <Setter Property="CornerRadius" Value="11" />
+    <Setter Property="CornerRadius" Value="10" />
+    <Setter Property="Foreground" Value="{StaticResource Text}" />
+    <Setter Property="FontFamily" Value="{StaticResource MonoFont}" />
+    <Setter Property="FontSize" Value="12" />
+  </Style>
+
+  <!--
+    ◆ The two port boxes, which are NumericUpDowns and were the only unthemed controls on this head.
+
+    They sit in a two-column Grid with a plain TextBox.field beside them — port next to username in the host
+    editor, and port next to default-user in the group editor — so the mismatch was not subtle: a bright
+    Fluent border and a near-black fill against this theme's #101019 and hairline, at the one place on the
+    screen where the two are a hand's width apart and directly comparable.
+
+    Not `Classes="field"` on the two call sites, because that class is a TextBox class and this is not a
+    TextBox: NumericUpDown is a ButtonSpinner wrapping one, so half the chrome being corrected belongs to
+    the spinner and would not take a TextBox setter. Typing the control instead also means a third port box
+    added later is right without anybody remembering, which is the property the desktop's own bare
+    `Selector="NumericUpDown"` in App.axaml has had since it was written.
+
+    Three selectors rather than one for the same reason Button.icon.accent needs two: Fluent puts the fill
+    and the border on template parts, and a value set on the control alone is a TemplateBinding the theme's
+    own wins over. The numbers are TextBox.field's, exactly — 44 tall, 10 radius, Field on BorderMid — so
+    the pair reads as one row of two boxes rather than as a box and a visitor.
+  -->
+  <Style Selector="NumericUpDown">
+    <Setter Property="Height" Value="44" />
+    <Setter Property="MinHeight" Value="44" />
+    <Setter Property="VerticalContentAlignment" Value="Center" />
+    <Setter Property="Background" Value="{StaticResource Field}" />
+    <Setter Property="BorderBrush" Value="{StaticResource BorderMid}" />
+    <Setter Property="BorderThickness" Value="1" />
+    <Setter Property="CornerRadius" Value="10" />
+    <Setter Property="Foreground" Value="{StaticResource Text}" />
+    <Setter Property="FontFamily" Value="{StaticResource MonoFont}" />
+    <Setter Property="FontSize" Value="12" />
+  </Style>
+  <Style Selector="NumericUpDown /template/ ButtonSpinner">
+    <Setter Property="Background" Value="{StaticResource Field}" />
+    <Setter Property="BorderBrush" Value="{StaticResource BorderMid}" />
+    <Setter Property="BorderThickness" Value="1" />
+    <Setter Property="CornerRadius" Value="10" />
+    <Setter Property="MinHeight" Value="44" />
+  </Style>
+  <!--
+    The inner box gives its chrome up entirely: the spinner above is drawing the border and the fill, and a
+    TextBox that kept its own would put a second rounded rectangle a pixel inside the first.
+  -->
+  <Style Selector="NumericUpDown /template/ TextBox">
+    <Setter Property="Background" Value="Transparent" />
+    <Setter Property="BorderThickness" Value="0" />
+    <Setter Property="Padding" Value="14,0" />
+    <Setter Property="VerticalContentAlignment" Value="Center" />
     <Setter Property="Foreground" Value="{StaticResource Text}" />
     <Setter Property="FontFamily" Value="{StaticResource MonoFont}" />
     <Setter Property="FontSize" Value="12" />
@@ -358,6 +602,31 @@
     <Setter Property="Fill" Value="{StaticResource Live}" />
   </Style>
 
+  <!--
+    Amber, and it does not contradict the remark above. That one says green is a fact about a host rather
+    than an accent, and this is the colour for a fact that is not settled yet: green is what is true, purple
+    is what you can press, and a connection still being made is neither. The palette's own rule gives amber
+    to the caveat worth reading, which is exactly what this is.
+
+    Only the tab strips use it, and only for a tab with no shell behind it yet — the same amber, from the
+    same brush, as the track and the running step on the connecting screen, so that a tab and the screen it
+    opens agree about what is happening. See TerminalScreen.axaml.
+  -->
+  <Style Selector="Ellipse.dot.connecting">
+    <Setter Property="Fill" Value="{StaticResource Warn}" />
+  </Style>
+
+  <!--
+    Red for a shell whose connection broke, splitting the grey that used to cover every dead tab. A shell
+    somebody exited and a shell that was cut off are the two dead states with the least in common, and on a
+    phone the difference matters more than it does on a desk: a mobile link drops on its own, and the tab
+    that went quiet while the screen was off is the one worth dialling again. See
+    TerminalTabViewModel.IsDropped.
+  -->
+  <Style Selector="Ellipse.dot.dropped">
+    <Setter Property="Fill" Value="{StaticResource Danger}" />
+  </Style>
+
   <!-- Every label, count, address and fingerprint in this design is monospace. See Palette.axaml. -->
   <Style Selector="TextBlock.mono">
     <Setter Property="FontFamily" Value="{StaticResource MonoFont}" />
@@ -370,10 +639,29 @@
     <Setter Property="FontSize" Value="9.5" />
     <Setter Property="FontWeight" Value="SemiBold" />
   </Style>
+  <!--
+    ◆ Wraps, like .body below it, and the omission was doing real damage rather than none.
+
+    A TextBlock that neither wraps nor trims does not shrink to fit and does not report that it did not:
+    it lays out at its full single-line width and the parent clips whatever runs past the edge. So a
+    .detail long enough to overflow ended mid-word against the screen, with no ellipsis to say anything
+    was missing — "A machine to connect to, filed under a group if you" on the add sheet, which reads as
+    a sentence somebody forgot to finish rather than as text that was cut.
+
+    It is a style rather than an attribute on the ones that overflow today, because which of the fifty-odd
+    .detail blocks overflows is a function of the device's width and the user's font scale, not of the
+    string: every one of them is a candidate at 320dp that is fine at 411dp, and fixing the handful that
+    happen to be visibly wrong on a Pixel 5 would leave the rest to be found one narrow phone at a time.
+
+    The nine that genuinely want one line — list rows and log lines, where a wrapped detail would push the
+    row's own height around — now say TextWrapping="NoWrap" beside their TextTrimming, which is the pairing
+    that was always meant and was previously being had by accident.
+  -->
   <Style Selector="TextBlock.detail">
     <Setter Property="FontFamily" Value="{StaticResource MonoFont}" />
     <Setter Property="Foreground" Value="{StaticResource TextFaint}" />
     <Setter Property="FontSize" Value="10" />
+    <Setter Property="TextWrapping" Value="Wrap" />
   </Style>
   <Style Selector="TextBlock.body">
     <Setter Property="FontFamily" Value="{StaticResource MonoFont}" />

src/DodoSSH.Client.Android/Views/FilesScreen.axaml

@@ -48,7 +48,12 @@
     the refusal has no continue button here either.
   -->
 
-  <Grid RowDefinitions="Auto,Auto,Auto,Auto,*,Auto">
+  <!--
+    Nine rows since the disconnect strip arrived; it is row 7, between the listing and the action bar, so
+    that what it says is next to the buttons it replaces the point of. Auto and empty except while the
+    session has been lost, so an ordinary browse pays nothing for it.
+  -->
+  <Grid RowDefinitions="Auto,Auto,Auto,Auto,Auto,Auto,*,Auto,Auto">
 
     <!-- ============ header ============ -->
     <Grid Grid.Row="0" ColumnDefinitions="Auto,Auto,*,Auto" Height="56" Margin="8,0">
@@ -78,7 +83,13 @@
     </Grid>
 
     <!-- ============ opening one ============ -->
-    <Border Grid.Row="1" Classes="card" Margin="12,0,12,8" IsVisible="{Binding !IsConnected}">
+    <!--
+      ShowsRemoteInvitation rather than !IsConnected, which stopped being the same question when a
+      connection could be lost rather than only closed: a session that went away leaves its listing on
+      screen, and this card springing back over it would hide the one thing the strip below is offering to
+      put the user back into. See TransfersViewModel.ShowsRemoteContents.
+    -->
+    <Border Grid.Row="1" Classes="card" Margin="12,0,12,8" IsVisible="{Binding ShowsRemoteInvitation}">
       <StackPanel Spacing="10">
 
         <ComboBox HorizontalAlignment="Stretch" MinHeight="44" IsVisible="{Binding ShowsHostPicker}"
@@ -174,11 +185,50 @@
       </StackPanel>
     </Border>
 
+    <!-- ============ ◆ typing where to go ============ -->
+    <!--
+      The breadcrumb below reaches only what is under the directory already listed, so crossing a host's
+      filesystem was a tap and a round trip a level and a path somebody already knew could not be entered at
+      all. This is the way in for one that is: absolute, relative to the directory showing, or starting at
+      ~. See SftpPath.Resolve.
+
+      GO is a real button rather than only the keyboard's own action key, and that is a phone decision: a
+      software keyboard's return key is whatever the IME decided to draw, this application does not set the
+      IME action, and a path box whose only commit was a key that may say "newline" would be a box nothing
+      could be submitted from. The Enter binding stays for the hardware keyboards this head does meet.
+    -->
+    <Grid Grid.Row="3" ColumnDefinitions="*,8,Auto" Margin="14,0,14,6"
+          IsVisible="{Binding ShowsRemoteContents}">
+      <TextBox Grid.Column="0" Classes="field" Text="{Binding RemotePathDraft}"
+               PlaceholderText="/path/to/somewhere">
+        <TextBox.KeyBindings>
+          <KeyBinding Gesture="Enter" Command="{Binding GoToTypedRemotePathCommand}" />
+        </TextBox.KeyBindings>
+      </TextBox>
+      <Button Grid.Column="2" Classes="secondary" Height="44" Padding="16,0" Content="GO"
+              Command="{Binding GoToTypedRemotePathCommand}" />
+    </Grid>
+
+    <!--
+      ◆ How many are gathered up, and the one gesture that gathers or releases them all. Only while
+      something is — an always-present row would spend 30 pixels of a listing on a count of zero.
+    -->
+    <Grid Grid.Row="3" ColumnDefinitions="Auto,*" Margin="14,0,14,6" VerticalAlignment="Bottom"
+          IsVisible="{Binding HasCheckedRemote}">
+      <Button Grid.Column="0" Classes="secondary" MinHeight="36" Padding="12,0"
+              Command="{Binding ToggleAllRemoteCommand}">
+        <TextBlock Classes="label" FontSize="9" Text="ALL / NONE" />
+      </Button>
+      <TextBlock Grid.Column="1" Classes="mono" FontSize="12" Margin="10,0,0,0"
+                 VerticalAlignment="Center" Foreground="{StaticResource AccentText}"
+                 Text="{Binding CheckedRemoteLabel}" />
+    </Grid>
+
     <!-- ============ where we are ============ -->
-    <ScrollViewer Grid.Row="3" HorizontalScrollBarVisibility="Auto" VerticalScrollBarVisibility="Disabled"
-                  IsVisible="{Binding IsConnected}" Margin="0,0,0,4">
+    <ScrollViewer Grid.Row="4" HorizontalScrollBarVisibility="Auto" VerticalScrollBarVisibility="Disabled"
+                  IsVisible="{Binding ShowsRemoteContents}" Margin="0,0,0,4">
       <StackPanel Orientation="Horizontal" Spacing="4" Margin="14,0" VerticalAlignment="Center">
-        <Button Classes="row" MinHeight="36" Padding="9,0" CornerRadius="9"
+        <Button Classes="row" MinHeight="36" Padding="9,0" CornerRadius="10"
                 Command="{Binding RemoteUpCommand}">
           <TextBlock Classes="mono" FontSize="12" Text="↑" Foreground="{StaticResource AccentText}" />
         </Button>
@@ -188,7 +238,7 @@
           </ItemsControl.ItemsPanel>
           <ItemsControl.ItemTemplate>
             <DataTemplate x:DataType="vm:CrumbViewModel">
-              <Button Classes="row" MinHeight="36" Padding="7,0" CornerRadius="9"
+              <Button Classes="row" MinHeight="36" Padding="7,0" CornerRadius="10"
                       Command="{Binding $parent[views:FilesScreen].((vm:TransfersViewModel)DataContext).GoRemoteCommand}"
                       CommandParameter="{Binding Path}">
                 <TextBlock Classes="detail" FontSize="11" Text="{Binding Name}" />
@@ -199,10 +249,50 @@
       </StackPanel>
     </ScrollViewer>
 
-    <!-- ============ the listing ============ -->
-    <Panel Grid.Row="4">
+    <!-- ============ pins ============ -->
+    <!--
+      ◆ The phone's answer to the desktop's QUICK ACCESS sidebar — the same PinnedPathList this host was
+      connected with, drawn where this head actually browses files rather than beside a terminal it has no
+      strip for. See TransfersViewModel.ConnectedPinnedPaths for why this is a snapshot taken at connect
+      rather than a live follow of the host row: a pin added or removed mid-session shows up here on the
+      next connect, not while this one is still open.
 
-      <TextBlock Classes="body" IsVisible="{Binding !IsConnected}" Margin="24,12"
+      Its own row rather than folded into the breadcrumb above: the two scroll independently and mean
+      different things — the breadcrumb is where this pane is, the chips are places it can jump to — and a
+      shared row would make ↑ look like one more pin among several.
+
+      Gone rather than empty when nothing is pinned or nothing is connected, which HasConnectedPins already
+      answers: an empty scrolling strip under the breadcrumb would read as a loading row rather than as
+      "this host has nothing pinned".
+    -->
+    <ScrollViewer Grid.Row="5" HorizontalScrollBarVisibility="Auto" VerticalScrollBarVisibility="Disabled"
+                  IsVisible="{Binding HasConnectedPins}" Margin="0,0,0,4">
+      <ItemsControl ItemsSource="{Binding ConnectedPinnedPaths}">
+        <ItemsControl.ItemsPanel>
+          <ItemsPanelTemplate><StackPanel Orientation="Horizontal" Spacing="6" Margin="14,0" /></ItemsPanelTemplate>
+        </ItemsControl.ItemsPanel>
+        <ItemsControl.ItemTemplate>
+          <DataTemplate x:DataType="x:String">
+            <!--
+              Classes="row" at its default 44dp MinHeight rather than the breadcrumb's own 36 — these chips
+              are the destination, not the trail behind it, and a target worth a deliberate tap gets the
+              full touch floor this head holds everything else to. Not trimmed: a chip scrolls sideways with
+              the row rather than being squeezed to fit it, so the full path is always what a tap commits to.
+            -->
+            <Button Classes="row" MinHeight="44" Padding="11,0" CornerRadius="10"
+                    Command="{Binding $parent[views:FilesScreen].((vm:TransfersViewModel)DataContext).GoRemoteCommand}"
+                    CommandParameter="{Binding}">
+              <TextBlock Classes="mono" FontSize="12" Text="{Binding}" />
+            </Button>
+          </DataTemplate>
+        </ItemsControl.ItemTemplate>
+      </ItemsControl>
+    </ScrollViewer>
+
+    <!-- ============ the listing ============ -->
+    <Panel Grid.Row="6">
+
+      <TextBlock Classes="body" IsVisible="{Binding ShowsRemoteInvitation}" Margin="24,12"
                  VerticalAlignment="Top" Text="{Binding Status}" />
 
       <!--
@@ -219,7 +309,7 @@
         selection navigate would send a single desktop click walking into a directory.
       -->
       <ListBox ItemsSource="{Binding RemoteEntries}" SelectedItem="{Binding SelectedRemoteEntry}"
-               Background="Transparent" BorderThickness="0" IsVisible="{Binding IsConnected}"
+               Background="Transparent" BorderThickness="0" IsVisible="{Binding ShowsRemoteContents}"
                Tapped="OnRemoteEntryTapped">
         <ListBox.Styles>
           <Style Selector="ListBoxItem">
@@ -236,15 +326,27 @@
         <ListBox.ItemTemplate>
           <DataTemplate x:DataType="vm:RemoteEntryRowViewModel">
             <Border MinHeight="52" Padding="10,7">
-              <Grid ColumnDefinitions="Auto,*,Auto">
-                <TextBlock Grid.Column="0" Width="16" HorizontalAlignment="Center"
+              <Grid ColumnDefinitions="Auto,Auto,*,Auto">
+                <!--
+                  ◆ The tick, which on this head is the only multi-selection there can be: a phone has no
+                  ctrl-click and no shift-click. It is not the list's selection — that stays "the row I am
+                  looking at", which RENAME, MOVE and PERMS act on one at a time — and this is "the rows I
+                  have gathered up", which SAVE and DELETE take. See RemoteEntryRowViewModel.IsChecked.
+
+                  Its own 40-pixel column so a thumb aimed at it does not open the directory beside it: the
+                  row's own tap navigates, which is this head's one behaviour the desktop does not share.
+                -->
+                <CheckBox Grid.Column="0" Width="40" MinWidth="0" VerticalAlignment="Center"
+                          IsChecked="{Binding IsChecked}" Click="OnRemoteRowChecked" />
+
+                <TextBlock Grid.Column="1" Width="16" HorizontalAlignment="Center"
                            VerticalAlignment="Center" FontSize="12" Text="▸"
                            Foreground="{StaticResource AccentText}" IsVisible="{Binding IsNavigable}" />
-                <TextBlock Grid.Column="0" Width="16" HorizontalAlignment="Center"
+                <TextBlock Grid.Column="1" Width="16" HorizontalAlignment="Center"
                            VerticalAlignment="Center" FontSize="12" Text="·"
                            Foreground="{StaticResource TextFaint}" IsVisible="{Binding IsFile}" />
 
-                <StackPanel Grid.Column="1" Margin="11,0,0,0" Spacing="2" VerticalAlignment="Center">
+                <StackPanel Grid.Column="2" Margin="11,0,0,0" Spacing="2" VerticalAlignment="Center">
                   <TextBlock Classes="mono" FontSize="12.5" Text="{Binding Name}"
                              TextTrimming="CharacterEllipsis" />
                   <StackPanel Orientation="Horizontal" Spacing="7">
@@ -255,7 +357,7 @@
                   </StackPanel>
                 </StackPanel>
 
-                <TextBlock Grid.Column="2" Classes="detail" Text="{Binding Modified}"
+                <TextBlock Grid.Column="3" Classes="detail" Text="{Binding Modified}"
                            VerticalAlignment="Center" />
               </Grid>
             </Border>
@@ -264,8 +366,50 @@
       </ListBox>
     </Panel>
 
+    <!-- ============ what the screen says once the session has stopped ============ -->
+    <!--
+      The phone's half of the desktop's SessionEndedBar, and it binds this screen's own view model rather
+      than the shell's — this control's DataContext is TransfersViewModel, where the desktop's strip sits in
+      MainWindow beside a status bar and can ask which surface it is on. The facts are the same two:
+      TransfersViewModel.LostConnection and the reconnect that goes back to the directory showing.
+
+      A phone needs it more than a desk does, and for the reason its terminal does: a mobile link drops on
+      its own and the screen goes off, so this session is discovered later with a listing on it and no
+      account of why nothing works. The listing stays deliberately — see ShowsRemoteContents.
+    -->
+    <Border Grid.Row="7" IsVisible="{Binding IsConnectionLost}" Background="{StaticResource DeepChrome}"
+            BorderBrush="{StaticResource Danger}" BorderThickness="0,1,0,0" Padding="12,10">
+      <StackPanel Spacing="8">
+
+        <StackPanel Orientation="Horizontal" Spacing="7" VerticalAlignment="Center">
+          <Ellipse Classes="dot" Classes.dropped="True" Width="7" Height="7" VerticalAlignment="Center" />
+          <TextBlock Classes="label" FontSize="9" Text="DISCONNECTED" />
+        </StackPanel>
+
+        <!--
+          Two lines maximum, for the reason the terminal's strip caps its own: a transport error can hand
+          over whatever the operating system called it, and a strip that grew to fit one would eat the
+          listing above it on a screen that has none to spare.
+        -->
+        <TextBlock Classes="detail" FontSize="10.5" Foreground="{StaticResource TextDim}"
+                   TextWrapping="Wrap" MaxLines="2" TextTrimming="CharacterEllipsis"
+                   Text="{Binding LostConnection}" />
+
+        <StackPanel Orientation="Horizontal" Spacing="8">
+          <Button Classes="row" MinHeight="44" Padding="14,0" Command="{Binding ReconnectCommand}">
+            <TextBlock Classes="label" FontSize="9" Text="RECONNECT" />
+          </Button>
+          <Button Classes="row" MinHeight="44" Padding="14,0" Command="{Binding DisconnectCommand}">
+            <TextBlock Classes="label" FontSize="9" Text="CLOSE" />
+          </Button>
+        </StackPanel>
+
+      </StackPanel>
+    </Border>
+
     <!-- ============ what to do with the chosen entry ============ -->
-    <Border Grid.Row="5" IsVisible="{Binding IsConnected}" Background="{StaticResource Chrome}"
+    <!-- DeepChrome rather than Chrome, since v5 — see the remark on PhoneShell's vault header. -->
+    <Border Grid.Row="8" IsVisible="{Binding IsConnected}" Background="{StaticResource DeepChrome}"
             BorderBrush="{StaticResource Border}" BorderThickness="0,1,0,0" Padding="12,10">
       <StackPanel Spacing="9">
 
@@ -287,6 +431,105 @@
           </StackPanel>
         </Border>
 
+        <!--
+          ◆ Renaming and moving, in the place the deletion question already occupies and never at the same
+          time as it. One form for both — see RemoteMoveRequest for why the two questions share a shape and
+          differ only in their words — and no danger colours on it, because a rename is undone by renaming
+          back and the deletion above is the one thing here that nothing can walk back.
+        -->
+        <Border IsVisible="{Binding IsMovingRemote}" Background="{StaticResource Raised}"
+                BorderBrush="{StaticResource Border}" BorderThickness="1" CornerRadius="12"
+                Padding="13,11">
+          <StackPanel Spacing="6">
+            <TextBlock Classes="mono" FontSize="12" TextWrapping="Wrap"
+                       Text="{Binding PendingRemoteMove.Question}" />
+            <TextBlock Classes="detail" TextWrapping="Wrap" Text="{Binding PendingRemoteMove.FullPath}" />
+            <TextBox Classes="field" Text="{Binding RemoteMoveTarget}"
+                     PlaceholderText="{Binding PendingRemoteMove.FieldHint}">
+              <TextBox.KeyBindings>
+                <KeyBinding Gesture="Enter" Command="{Binding ConfirmMoveRemoteCommand}" />
+              </TextBox.KeyBindings>
+            </TextBox>
+            <TextBlock Classes="body" Text="{Binding PendingRemoteMove.Note}" />
+            <Grid ColumnDefinitions="*,8,*" Margin="0,4,0,0">
+              <Button Grid.Column="0" Classes="primary" Height="44"
+                      Content="{Binding PendingRemoteMove.ConfirmLabel}"
+                      Command="{Binding ConfirmMoveRemoteCommand}" IsEnabled="{Binding !IsBusy}" />
+              <Button Grid.Column="2" Classes="secondary" Height="44" Content="CANCEL"
+                      Command="{Binding CancelMoveRemoteCommand}" />
+            </Grid>
+          </StackPanel>
+        </Border>
+
+        <!--
+          ◆ The mode, as nine boxes and the number they add up to — in the place the other two questions
+          already occupy and never at the same time as either. See RemotePermissionsRequest for why it is
+          boxes rather than a field to type 755 into, and TransfersScreen.axaml for the desktop's own copy
+          of the same form.
+
+          Compact rather than a second layout: the three columns are the mode's own three triples, and at
+          360dp they are exactly what fits. The octal is printed underneath as it changes, which is what
+          lets somebody who came here to type chmod check the boxes they just ticked.
+        -->
+        <Border IsVisible="{Binding IsChangingRemotePermissions}" Background="{StaticResource Raised}"
+                BorderBrush="{StaticResource Border}" BorderThickness="1" CornerRadius="12"
+                Padding="13,11">
+          <StackPanel Spacing="6">
+            <TextBlock Classes="mono" FontSize="12" TextWrapping="Wrap"
+                       Text="{Binding PendingRemotePermissions.Question}" />
+            <TextBlock Classes="detail" TextWrapping="Wrap"
+                       Text="{Binding PendingRemotePermissions.FullPath}" />
+
+            <Grid ColumnDefinitions="Auto,*,*,*" RowDefinitions="Auto,44,44,44" Margin="0,2">
+              <TextBlock Grid.Row="0" Grid.Column="1" Classes="detail" Text="read"
+                         HorizontalAlignment="Center" />
+              <TextBlock Grid.Row="0" Grid.Column="2" Classes="detail" Text="write"
+                         HorizontalAlignment="Center" />
+              <TextBlock Grid.Row="0" Grid.Column="3" Classes="detail"
+                         Text="{Binding PendingRemotePermissions.ExecuteLabel}"
+                         HorizontalAlignment="Center" />
+
+              <TextBlock Grid.Row="1" Grid.Column="0" Classes="mono" FontSize="12" Text="owner"
+                         VerticalAlignment="Center" Margin="0,0,10,0" />
+              <CheckBox Grid.Row="1" Grid.Column="1" HorizontalAlignment="Center" MinWidth="0"
+                        IsChecked="{Binding PendingRemotePermissions.OwnerRead}" />
+              <CheckBox Grid.Row="1" Grid.Column="2" HorizontalAlignment="Center" MinWidth="0"
+                        IsChecked="{Binding PendingRemotePermissions.OwnerWrite}" />
+              <CheckBox Grid.Row="1" Grid.Column="3" HorizontalAlignment="Center" MinWidth="0"
+                        IsChecked="{Binding PendingRemotePermissions.OwnerExecute}" />
+
+              <TextBlock Grid.Row="2" Grid.Column="0" Classes="mono" FontSize="12" Text="group"
+                         VerticalAlignment="Center" Margin="0,0,10,0" />
+              <CheckBox Grid.Row="2" Grid.Column="1" HorizontalAlignment="Center" MinWidth="0"
+                        IsChecked="{Binding PendingRemotePermissions.GroupRead}" />
+              <CheckBox Grid.Row="2" Grid.Column="2" HorizontalAlignment="Center" MinWidth="0"
+                        IsChecked="{Binding PendingRemotePermissions.GroupWrite}" />
+              <CheckBox Grid.Row="2" Grid.Column="3" HorizontalAlignment="Center" MinWidth="0"
+                        IsChecked="{Binding PendingRemotePermissions.GroupExecute}" />
+
+              <TextBlock Grid.Row="3" Grid.Column="0" Classes="mono" FontSize="12" Text="others"
+                         VerticalAlignment="Center" Margin="0,0,10,0" />
+              <CheckBox Grid.Row="3" Grid.Column="1" HorizontalAlignment="Center" MinWidth="0"
+                        IsChecked="{Binding PendingRemotePermissions.OthersRead}" />
+              <CheckBox Grid.Row="3" Grid.Column="2" HorizontalAlignment="Center" MinWidth="0"
+                        IsChecked="{Binding PendingRemotePermissions.OthersWrite}" />
+              <CheckBox Grid.Row="3" Grid.Column="3" HorizontalAlignment="Center" MinWidth="0"
+                        IsChecked="{Binding PendingRemotePermissions.OthersExecute}" />
+            </Grid>
+
+            <TextBlock Classes="mono" FontSize="13" Foreground="{StaticResource AccentText}"
+                       Text="{Binding PendingRemotePermissions.Preview}" />
+
+            <Grid ColumnDefinitions="*,8,*" Margin="0,4,0,0">
+              <Button Grid.Column="0" Classes="primary" Height="44" Content="APPLY"
+                      Command="{Binding ConfirmRemotePermissionsCommand}"
+                      IsEnabled="{Binding PendingRemotePermissions.HasChanges}" />
+              <Button Grid.Column="2" Classes="secondary" Height="44" Content="CANCEL"
+                      Command="{Binding CancelRemotePermissionsCommand}" />
+            </Grid>
+          </StackPanel>
+        </Border>
+
         <!--
           ◆ The queue, above the actions and only when it has something in it. Bounded and scrolling rather
           than growing: five files queued must not push the buttons off the bottom of the screen, which on a
@@ -308,7 +551,7 @@
                                  Foreground="{StaticResource Accent}"
                                  Background="{StaticResource Raised}" />
                     <TextBlock Classes="detail" Text="{Binding Progress}"
-                               TextTrimming="CharacterEllipsis" />
+                               TextWrapping="NoWrap" TextTrimming="CharacterEllipsis" />
                   </StackPanel>
 
                   <!--
@@ -345,20 +588,44 @@
           save picker names one destination. CanDownload is the desktop's own flag for the same question —
           a file is selected and something is connected — and is reused rather than restated here.
         -->
-        <Grid ColumnDefinitions="*,8,*" IsVisible="{Binding !IsConfirmingRemoteDeletion}">
+        <Grid ColumnDefinitions="*,8,*" IsVisible="{Binding !IsAskingAboutRemoteEntry}">
           <Button Grid.Column="0" Classes="primary" Height="44" Content="ADD FILES" Click="OnAddFiles" />
           <Button Grid.Column="2" Classes="secondary" Height="44" Content="SAVE FILE" Click="OnSaveFile"
                   IsEnabled="{Binding CanDownload}" />
         </Grid>
 
-        <Grid ColumnDefinitions="*,8,*" IsVisible="{Binding !IsConfirmingRemoteDeletion}">
-          <Button Grid.Column="0" Classes="secondary" Height="44" Content="DELETE"
-                  Command="{Binding DeleteRemoteCommand}" IsEnabled="{Binding CanDeleteRemote}" />
-          <Button Grid.Column="2" Classes="secondary" Height="44" Content="CLOSE"
-                  Command="{Binding DisconnectCommand}" />
+        <!--
+          ◆ The two that put a file right rather than taking it away, above the row DELETE is on and away
+          from it. Both open a form in place of these buttons — see the card above — because a phone has
+          nowhere to put a text field except where something else already was.
+        -->
+        <Grid ColumnDefinitions="*,8,*" IsVisible="{Binding !IsAskingAboutRemoteEntry}">
+          <Button Grid.Column="0" Classes="secondary" Height="44" Content="RENAME"
+                  Command="{Binding RenameRemoteCommand}" IsEnabled="{Binding CanRenameRemote}" />
+          <Button Grid.Column="2" Classes="secondary" Height="44" Content="MOVE"
+                  Command="{Binding MoveRemoteCommand}" IsEnabled="{Binding CanMoveRemote}" />
         </Grid>
 
-        <TextBlock Classes="body" IsVisible="{Binding !IsConfirmingRemoteDeletion}"
+        <!--
+          ◆ PERMS beside DELETE rather than on the row above, which is full: it is the fourth thing that can
+          be done to a chosen row and the third that puts one right. Dark on a bucket, where an object has an
+          access policy rather than nine bits — see CanChangeRemotePermissions.
+        -->
+        <Grid ColumnDefinitions="*,8,*" IsVisible="{Binding !IsAskingAboutRemoteEntry}">
+          <Button Grid.Column="0" Classes="secondary" Height="44" Content="PERMS"
+                  Command="{Binding ChangeRemotePermissionsCommand}"
+                  IsEnabled="{Binding CanChangeRemotePermissions}" />
+          <Button Grid.Column="2" Classes="secondary" Height="44" Content="DELETE"
+                  Command="{Binding DeleteRemoteCommand}" IsEnabled="{Binding CanDeleteRemote}" />
+        </Grid>
+
+        <Grid ColumnDefinitions="*,8,*" IsVisible="{Binding !IsAskingAboutRemoteEntry}">
+          <Button Grid.Column="0" Classes="secondary" Height="44" Content="CLOSE"
+                  Command="{Binding DisconnectCommand}" />
+          <Panel Grid.Column="2" />
+        </Grid>
+
+        <TextBlock Classes="body" IsVisible="{Binding !IsAskingAboutRemoteEntry}"
                    Text="ADD FILES picks documents to send. SAVE FILE asks where the selected file should be kept — the file is created there when you choose it, so a transfer that fails leaves it empty." />
 
         <TextBlock Classes="detail" Foreground="{StaticResource TextDim}" TextWrapping="Wrap"

src/DodoSSH.Client.Android/Views/FilesScreen.axaml.cs

@@ -35,6 +35,15 @@ internal sealed partial class FilesScreen : UserControl
     /// read its size the same gesture as fetching it.
     /// </para>
     /// </remarks>
+    /// <summary>Tells the screen a row's tick moved, so the count above the listing keeps up.</summary>
+    /// <remarks>
+    /// The desktop's own handler, for the reason given there: the count is a property of the collection
+    /// rather than of any row, and each listing replaces every row it has. See
+    /// <c>TransfersViewModel.RemoteChecksChanged</c>.
+    /// </remarks>
+    private void OnRemoteRowChecked(object? sender, RoutedEventArgs e) =>
+        (DataContext as TransfersViewModel)?.RemoteChecksChanged();
+
     private void OnRemoteEntryTapped(object? sender, TappedEventArgs e)
     {
         if (DataContext is TransfersViewModel { SelectedRemoteEntry.IsNavigable: true } transfers)

src/DodoSSH.Client.Android/Views/HostActionBar.axaml

@@ -0,0 +1,65 @@
+<UserControl xmlns="https://github.com/avaloniaui"
+             xmlns:x="http://schemas.microsoft.com/winfx/2006/xaml"
+             xmlns:vm="using:DodoSSH.Client.Shell.ViewModels"
+             x:Class="DodoSSH.Client.Android.Views.HostActionBar"
+             x:DataType="vm:VaultViewModel">
+
+  <!--
+    ◆ THE CONTEXTUAL ACTION BAR, WHICH IS WHAT THE CONNECT CARD BECAME.
+
+    A long press on a host used to raise a card over the bottom of the list: a password box, CONNECT, EDIT,
+    MOVE and DELETE. It was a menu drawn as a form, in the one part of the screen a list grows into — so it
+    covered rows, it had room for five things and never a sixth, and everything on it was about exactly one
+    machine. What replaced it is a selection and this bar.
+
+    It takes the vault header's place rather than sitting under it, which is the arrangement Android has used
+    for this since contextual action bars existed and is the reason it can be unambiguous: while it is up, the
+    screen is about the ticked hosts and about nothing else. See PhoneShell.ShowsHostSelectionBar, which is
+    where the swap happens and why it is computed rather than bound.
+
+    Left to right: the cross that leaves selection mode, the count, then the two controls. The pencil is out
+    in front of the menu because editing is the thing people reach for most and is worth not opening a menu
+    for; the other six are behind the ⋯, which is a sheet rather than a flyout — see
+    VaultViewModel.IsHostActionSheetOpen.
+
+    ── one host or several ──────────────────────────────────────────────────────────────────────────────────
+    The pencil is drawn only while exactly one host is ticked, and it is collapsed rather than greyed. Three
+    of the seven actions are like that — edit, connect, and browse — and the reason is the same for all
+    three: a form, a terminal and a file session are each about one machine, and there is no sensible reading
+    of "edit these six". The three that stay are the ones a count makes better rather than worse.
+  -->
+
+  <!-- DeepChrome rather than Chrome, since v5: this bar takes the vault header's own place, so it takes
+       the vault header's own surface too — see the remark on that header in PhoneShell.axaml. -->
+  <Border Background="{StaticResource DeepChrome}" BorderBrush="{StaticResource Border}"
+          BorderThickness="0,0,0,1" Padding="6,0" Height="56">
+    <Grid ColumnDefinitions="Auto,*,Auto,Auto">
+
+      <!--
+        The cross, and it is the first thing on the bar for the reason back arrows are: the way out of a
+        mode belongs at the edge the thumb reaches without crossing anything it might press by mistake.
+      -->
+      <Button Grid.Column="0" Classes="icon" Content="✕"
+              Command="{Binding ClearHostChoiceCommand}"
+              ToolTip.Tip="Stop choosing hosts" />
+
+      <!--
+        The count, and only the count. The bar it sits in is already saying what the number is about, and
+        "6 hosts selected" beside a pencil and a menu is the width those two icons need at 360dp.
+      -->
+      <TextBlock Grid.Column="1" Classes="heading" FontSize="17" Margin="8,0"
+                 VerticalAlignment="Center" Text="{Binding ChosenHostsLabel}" />
+
+      <Button Grid.Column="2" Classes="icon" Content="✎" FontSize="15"
+              IsVisible="{Binding HasOneChosenHost}"
+              Command="{Binding EditChosenHostCommand}"
+              ToolTip.Tip="Edit this host" />
+
+      <Button Grid.Column="3" Classes="icon" Content="⋯" FontSize="18"
+              Command="{Binding OpenHostActionSheetCommand}"
+              ToolTip.Tip="More things to do with these hosts" />
+
+    </Grid>
+  </Border>
+
+</UserControl>

src/DodoSSH.Client.Android/Views/HostActionBar.axaml.cs

@@ -0,0 +1,10 @@
+using Avalonia.Controls;
+using Avalonia.Markup.Xaml;
+
+namespace DodoSSH.Client.Android.Views;
+
+/// <summary>The bar that takes the header's place while hosts are chosen.</summary>
+internal sealed partial class HostActionBar : UserControl
+{
+    public HostActionBar() => AvaloniaXamlLoader.Load(this);
+}

src/DodoSSH.Client.Android/Views/HostKeySheet.axaml

@@ -17,6 +17,13 @@
     and the plan names presenting it as dismissible as the one design mistake that matters here. A phone
     makes this easy to get wrong, because a bottom sheet is swipe-to-dismiss by convention; this is not a
     sheet for that reason, it is a full-screen panel.
+
+    ◆ WHAT IS BEHIND IT IS NO LONGER ALWAYS THE HOST LIST. The shell used to move to HOSTS before letting a
+    handshake ask this, because the desktop head drew the question as a banner on that screen; both heads draw
+    it over the surface now, so this is raised over whatever the user was on — including the connect box on the
+    Connections surface, which is where a machine that is in no keychain at all is dialled from. Nothing here
+    changes for that, and that is the point: this control never knew which screen it was over. See
+    MainWindowViewModel.OnVaultConnectionFailed and HostKeyCard.axaml on the desktop.
   -->
 
   <Panel>
@@ -29,7 +36,7 @@
         written out because the palette holds no alpha variant of a surface — see QuickConnect on the
         desktop, which carries the same note.
       -->
-      <Border Background="#9E0E1220" />
+      <Border Background="#9E05050A" />
 
       <Border VerticalAlignment="Bottom" Background="{StaticResource Panel}"
               BorderBrush="{StaticResource BorderMid}" BorderThickness="0,1,0,0"
@@ -42,12 +49,29 @@
 
             <TextBlock Classes="title" Text="UNKNOWN HOST KEY" FontSize="13" />
 
+            <!--
+              ◆ The port and the full stop are one Run, and that is the whole of the fix.
+
+              Avalonia puts a space between adjacent Runs, which is what makes "First contact with", the
+              host and the separator read as a sentence without any of them carrying padding. The trailing
+              Run began with a full stop, so it got that space too and the sheet said
+
+                  First contact with 10.0.2.2 · 2222 . Nothing in this keychain…
+
+              — a gap before the stop, on the one dialogue in this application whose entire job is to be
+              read carefully before somebody answers it. StringFormat welds the stop to the number inside a
+              single Run, where no inter-Run space can get in, and the sentence that follows starts a Run of
+              its own so it still gets the space it wants.
+
+              The `{}` prefix is required: a StringFormat beginning with a brace would otherwise be parsed
+              as a markup extension.
+            -->
             <TextBlock Classes="body" Margin="0,6,0,0">
               <Run Text="First contact with" />
               <Run Text="{Binding PendingHostKey.Host}" Foreground="{StaticResource Text}" />
               <Run Text="·" />
-              <Run Text="{Binding PendingHostKey.Port}" />
-              <Run Text=". Nothing in this keychain has approved this key." />
+              <Run Text="{Binding PendingHostKey.Port, StringFormat='{}{0}.'}" />
+              <Run Text="Nothing in this keychain has approved this key." />
             </TextBlock>
 
             <TextBlock Classes="label" Margin="0,14,0,5" Foreground="{StaticResource TextFaint}"
@@ -107,8 +131,15 @@
           <!--
             The only control on the screen, and it goes back rather than forward. Compare the sheet above,
             where the accent button connects.
+
+            ◆ IT SAID "BACK TO HOSTS", AND IT NEITHER WENT THERE NOR CAME BACK. Two things were wrong with
+            it at once. RejectHostKey cleared only the pending key, so pressing it left HasHostKeyMismatch
+            set and this opaque panel up over whatever the user navigated to next — including the host editor
+            the paragraph above sends them to. And the shell no longer moves to Hosts to ask the question, so
+            even fixed it does not land there: it lowers this panel and reveals the screen that was already
+            underneath. "BACK" is what that is. See VaultViewModel.RejectHostKey.
           -->
-          <Button Classes="secondary" Content="BACK TO HOSTS" Margin="0,24,0,0"
+          <Button Classes="secondary" Content="BACK" Margin="0,24,0,0"
                   Command="{Binding RejectHostKeyCommand}" />
 
         </StackPanel>

src/DodoSSH.Client.Android/Views/HostsScreen.axaml.cs

@@ -1,23 +1,26 @@
+using Avalonia;
 using Avalonia.Controls;
 using Avalonia.Input;
+using Avalonia.Interactivity;
 using Avalonia.Markup.Xaml;
+using Avalonia.VisualTree;
 
 using DodoSSH.Client.Shell.ViewModels;
 
 namespace DodoSSH.Client.Android.Views;
 
-/// <summary>Design 02 — the host list, and the connect bar that replaces the desktop's right column.</summary>
+/// <summary>Design 02 — the host list, and the two gestures that are the whole of this file.</summary>
 /// <remarks>
 /// <para>
-/// ◆ <b>Two gestures on one list, and this file is the whole of the difference between them.</b> A tap on a
-/// host connects to it; a long press asks about it, which is what raises the bar. Why they were split is on
-/// the screen itself; what is here is the mechanics, and there are two of them worth knowing.
+/// ◆ <b>A tap connects, a long press chooses, and once anything is chosen a tap ticks instead.</b> The
+/// branch in the middle of that is why the tap is handled here rather than bound in the markup: which of
+/// the two a tap means depends on whether the screen is in selection mode, and Avalonia's bindings cannot
+/// ask.
 /// </para>
 /// <para>
-/// The handlers are on the <c>ListBox</c> rather than on the row, and the row stays a plain <c>Grid</c>.
-/// A button as the item template swallows the press before the list sees it — <c>FilesScreen</c> writes that
-/// out at length — leaving nothing selected and every control that reads the selection doing nothing. Both
-/// events fire after the list has moved its selection, which is what lets these read it.
+/// The handlers are on the <c>ListBox</c> rather than on the row, and the row stays a plain
+/// <c>Border</c>. A button as the item template swallows the press before the list sees it —
+/// <c>FilesScreen</c> writes that out at length — so neither handler would ever run.
 /// </para>
 /// </remarks>
 internal sealed partial class HostsScreen : UserControl
@@ -27,18 +30,18 @@ internal sealed partial class HostsScreen : UserControl
     /// </summary>
     /// <remarks>
     /// Avalonia raises <c>Tapped</c> on release whatever the press lasted, so without this a long press
-    /// would open the bar and then connect — the two gestures firing one after the other on one touch, which
-    /// is the one outcome that would make both of them untrustworthy. Set when the hold starts and cleared
-    /// by the tap it suppresses, so it never survives the gesture that set it.
+    /// would tick the row and then the tap it becomes would untick it again — one touch doing a thing and
+    /// undoing it, which is the one outcome that would make both gestures untrustworthy. Set when the hold
+    /// starts and cleared by the tap it suppresses, so it never survives the gesture that set it.
     /// </remarks>
     private bool held;
 
     /// <remarks>
     /// <para>
-    /// The long press is attached here rather than in the markup so that it sits beside the property that
-    /// makes it fire at all. <see cref="InputElement.IsHoldingEnabledProperty"/> is set rather than assumed:
-    /// it is the whole of the gesture, and a default that changed would take it away silently — every tap
-    /// would go on working and nothing would ever open the bar again.
+    /// Both gestures are attached here rather than in the markup so that they sit beside the property that
+    /// makes one of them fire at all. <see cref="InputElement.IsHoldingEnabledProperty"/> is set rather than
+    /// assumed: it is the whole of the long press, and a default that changed would take it away silently —
+    /// every tap would go on working and nothing would ever choose a host again.
     /// </para>
     /// <para>
     /// ◆ <b><c>FindControl</c> rather than the field the name generator declares for <c>x:Name</c></b>,
@@ -62,20 +65,20 @@ internal sealed partial class HostsScreen : UserControl
     }
 
     /// <summary>
-    /// Connects to the row that was tapped.
+    /// Connects to the row that was tapped, or ticks it where hosts are already ticked.
     /// </summary>
     /// <remarks>
     /// <para>
-    /// Guarded on the selection being a host rather than on what was under the finger. A tap on a group
-    /// heading moves the list's selection and the view model bounces it straight back to whatever was chosen
-    /// before — see <c>VaultViewModel.SelectedSidebarRow</c> — so reading the selection here answers "a host,
-    /// or nothing" without this file needing to know that rule. The cost of getting it wrong is connecting to
-    /// a machine the user was not pointing at.
+    /// ◆ <b>The branch is the selection, not the row.</b> Once anything is chosen the screen is in selection
+    /// mode — the bar across the top says so — and in that mode a tap adds and removes rather than
+    /// connecting. That is what every Android list does, and the alternative is worse than a rule to
+    /// remember: a tap that connected while five machines sat ticked would open a terminal on top of a
+    /// selection somebody was halfway through building.
     /// </para>
     /// <para>
-    /// Fire-and-forget, as the desktop grid's activation is: the command reports its own failures onto the
-    /// status line — an unknown host key, a refused password — and awaiting it here would be an event handler
-    /// returning a task nothing observes.
+    /// Fire-and-forget where it connects, as the desktop grid's activation is: the command reports its own
+    /// failures onto the status line — an unknown host key, a refused password — and awaiting it here would
+    /// be an event handler returning a task nothing observes.
     /// </para>
     /// </remarks>
     private void OnRowTapped(object? sender, TappedEventArgs e)
@@ -86,25 +89,25 @@ internal sealed partial class HostsScreen : UserControl
             return;
         }
 
-        if (DataContext is VaultViewModel { SelectedHost: { } row } vault)
+        if (DataContext is not VaultViewModel vault || RowUnder(e) is not { } row)
         {
-            _ = vault.ConnectToRowCommand.ExecuteAsync(row);
+            return;
         }
+
+        if (vault.IsChoosingHosts)
+        {
+            vault.ToggleHostChoiceCommand.Execute(row);
+            return;
+        }
+
+        _ = vault.ConnectToRowCommand.ExecuteAsync(row);
     }
 
-    /// <summary>
-    /// Raises the bar about the row that was held.
-    /// </summary>
+    /// <summary>Puts a tick against the row that was held, entering selection mode with it.</summary>
     /// <remarks>
-    /// <para>
     /// On <see cref="HoldingState.Started"/> rather than on completion, so the bar is up while the finger is
     /// still down. A long press that showed nothing until release would be a gesture with no way to tell it
     /// had been recognised, and the only feedback available on this list is the thing it does.
-    /// </para>
-    /// <para>
-    /// It goes through <c>OpenHostPaneCommand</c>, which is the desktop's own "ask about this host" — the two
-    /// heads raise different furniture from one flag rather than keeping a selection rule each.
-    /// </para>
     /// </remarks>
     private void OnRowHeld(object? sender, HoldingRoutedEventArgs e)
     {
@@ -115,9 +118,33 @@ internal sealed partial class HostsScreen : UserControl
 
         held = true;
 
-        if (DataContext is VaultViewModel { SelectedHost: { } row } vault)
+        if (DataContext is VaultViewModel vault && RowUnder(e) is { } row)
         {
-            vault.OpenHostPaneCommand.Execute(row);
+            vault.ChooseHostCommand.Execute(row);
         }
     }
+
+    /// <summary>
+    /// The host the gesture landed on, or null where it landed on something that is not one.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// ◆ <b>Read off the element under the finger rather than off the list's selection, which is what both
+    /// handlers used to do.</b> The selection was defensible while a tap only ever connected: a tap on a
+    /// group heading moves the selection and the view model bounces it straight back — see
+    /// <c>VaultViewModel.SelectedSidebarRow</c> — so reading it answered "a host, or nothing" for free.
+    /// It stops being defensible the moment a tap can tick one: a heading would then bounce the selection to
+    /// whichever host was last chosen and tick <em>that</em>, which is a machine the user was not pointing
+    /// at going into a set they are about to delete.
+    /// </para>
+    /// <para>
+    /// The ancestor rather than <c>e.Source</c> itself, because the source is whatever leaf the finger
+    /// landed on — a tag chip's <c>TextBlock</c> has a string for a data context, and the row's own
+    /// <c>Ellipse</c> has the row. The <see cref="ListBoxItem"/> is the one element in that chain whose data
+    /// context is always the list's item, whatever kind it is.
+    /// </para>
+    /// </remarks>
+    private static HostRowViewModel? RowUnder(RoutedEventArgs e) =>
+        (e.Source as Visual)?.FindAncestorOfType<ListBoxItem>(includeSelf: true)?.DataContext
+            as HostRowViewModel;
 }

src/DodoSSH.Client.Android/Views/KeychainScreen.axaml

@@ -149,8 +149,19 @@
           </StackPanel>
         </StackPanel>
 
-        <Button Grid.Column="2" Classes="danger" Height="44" Width="104" Content="DELETE"
-                Command="{Binding DeleteSelectedItemCommand}" />
+        <!--
+          MOVE beside it, and only where there is somewhere to move to — the rule the host's MOVE follows on
+          this head, for the reason a phone has: there is no room to draw a button that answers with a
+          refusal. It is the ghost of the pair rather than the danger one, because a move is undone by
+          moving it back.
+        -->
+        <StackPanel Grid.Column="2" Orientation="Horizontal" Spacing="8">
+          <Button Classes="secondary" Height="44" Width="86" Content="MOVE"
+                  IsVisible="{Binding CanMoveSelectedItem}"
+                  Command="{Binding MoveSelectedItemCommand}" />
+          <Button Classes="danger" Height="44" Width="104" Content="DELETE"
+                  Command="{Binding DeleteSelectedItemCommand}" />
+        </StackPanel>
       </Grid>
     </Border>
 
@@ -180,6 +191,47 @@
       </StackPanel>
     </Border>
 
+    <!-- ============ ◆ moving it to another vault ============ -->
+    <!--
+      The desktop's panel, in the place the deletion question uses and never at the same time as it: MOVE
+      disarms a pending deletion on the way in, and the buttons that ask either question are hidden while
+      one is up.
+
+      Both sentences are here rather than only in the status line afterwards, which on a phone is one line at
+      the bottom of a screen somebody has already navigated away from. The second one is the count of what
+      points at this key — every one of them is re-aimed at the vault it moves to, and that is the part
+      nobody can see from a keychain row.
+    -->
+    <Border Grid.Row="3" IsVisible="{Binding IsMovingItem}" Margin="12,4"
+            Background="{StaticResource Panel}" BorderBrush="{StaticResource BorderMid}"
+            BorderThickness="1" CornerRadius="12" Padding="14,12">
+      <StackPanel Spacing="8">
+        <TextBlock Classes="label" Text="MOVE TO VAULT" />
+        <TextBlock Classes="mono" FontSize="12" TextWrapping="Wrap"
+                   Text="{Binding MovingItemLabel}" />
+        <ComboBox HorizontalAlignment="Stretch" MinHeight="44"
+                  ItemsSource="{Binding MoveItemVaultChoices}"
+                  SelectedItem="{Binding SelectedMoveItemVault}">
+          <ComboBox.ItemTemplate>
+            <DataTemplate x:DataType="vm:VaultChoiceViewModel">
+              <TextBlock Classes="mono" FontSize="12" Text="{Binding Display}" />
+            </DataTemplate>
+          </ComboBox.ItemTemplate>
+        </ComboBox>
+        <TextBlock Classes="body" TextWrapping="Wrap"
+                   Text="It is re-encrypted with the other vault's key, so everybody who holds that key can read it and nobody in the vault it leaves can." />
+        <TextBlock Classes="body" TextWrapping="Wrap"
+                   IsVisible="{Binding HasMovingItemUsage}"
+                   Text="{Binding MovingItemUsage}" />
+        <Grid ColumnDefinitions="*,8,*" Margin="0,4,0,0">
+          <Button Grid.Column="0" Classes="primary" Height="44" Content="MOVE"
+                  Command="{Binding ConfirmMoveItemCommand}" IsEnabled="{Binding !IsBusy}" />
+          <Button Grid.Column="2" Classes="secondary" Height="44" Content="CANCEL"
+                  Command="{Binding CancelMoveItemCommand}" />
+        </Grid>
+      </StackPanel>
+    </Border>
+
     <!-- ============ the items ============ -->
     <Panel Grid.Row="4">
 
@@ -235,7 +287,7 @@
                     </Border>
                   </StackPanel>
                   <!-- What is known *about* an item, never the secret. -->
-                  <TextBlock Classes="detail" Text="{Binding Detail}" TextTrimming="CharacterEllipsis" />
+                  <TextBlock Classes="detail" Text="{Binding Detail}" TextWrapping="NoWrap" TextTrimming="CharacterEllipsis" />
                 </StackPanel>
 
                 <TextBlock Grid.Column="2" Text="›" Foreground="{StaticResource TextGhost}" FontSize="15"

src/DodoSSH.Client.Android/Views/LockedScreen.axaml

@@ -91,8 +91,8 @@
 
           Not Border.card, and neither are its two counterparts on HOSTS and FILES: that class is a panel,
           and this is a warning, so the background and the border are the warn pair rather than the chrome
-          one. What it does take is the radius the v2 ladder gives anything card-sized, which is what the
-          other two already draw.
+          one. What it does take is the radius the ladder gives anything card-sized — 12, unchanged from v2
+          to v5 — which is what the other two already draw.
         -->
         <Border IsVisible="{Binding HasLiveSessions}" Margin="0,22,0,0"
                 Background="{StaticResource WarnWash}" BorderBrush="{StaticResource WarnSoft}"

src/DodoSSH.Client.Android/Views/LogsScreen.axaml

@@ -106,11 +106,11 @@
 
                     <TextBlock Classes="detail" FontSize="10" Text="{Binding HostLabel}"
                                Foreground="{StaticResource TextDim}" VerticalAlignment="Center"
-                               TextTrimming="CharacterEllipsis" />
+                               TextWrapping="NoWrap" TextTrimming="CharacterEllipsis" />
                   </StackPanel>
 
                   <TextBlock Classes="detail" FontSize="10.5" Foreground="{StaticResource Text}"
-                             Text="{Binding Address}" TextTrimming="CharacterEllipsis" />
+                             Text="{Binding Address}" TextWrapping="NoWrap" TextTrimming="CharacterEllipsis" />
 
                   <!--
                     TextDim rather than the `detail` class's own faint step. This block is drawn on
@@ -124,7 +124,7 @@
                     <TextBlock Classes="detail" FontSize="9.5" Foreground="{StaticResource TextDim}"
                                Text="·" />
                     <TextBlock Classes="detail" FontSize="9.5" Foreground="{StaticResource TextDim}"
-                               Text="{Binding DeviceName}" TextTrimming="CharacterEllipsis" />
+                               Text="{Binding DeviceName}" TextWrapping="NoWrap" TextTrimming="CharacterEllipsis" />
                   </StackPanel>
 
                   <!-- A refused host key is the one outcome worth a colour of its own. -->
@@ -158,7 +158,7 @@
                   </StackPanel>
 
                   <TextBlock Classes="detail" FontSize="10.5" Foreground="{StaticResource Text}"
-                             Text="{Binding ItemLabel}" TextTrimming="CharacterEllipsis" />
+                             Text="{Binding ItemLabel}" TextWrapping="NoWrap" TextTrimming="CharacterEllipsis" />
 
                   <!-- The names of the fields that changed, never their values. -->
                   <TextBlock Classes="detail" FontSize="9.5" IsVisible="{Binding HasChangedFields}"

src/DodoSSH.Client.Android/Views/PhoneRail.axaml

@@ -36,6 +36,18 @@
     The .railentry styles live in this file because they are this control's own shape and nothing else wears
     them; every colour in them is resolved from Theme/Palette.axaml by key, which is the rule that matters —
     a rail that named its own blues is how a fifth surface colour ends up in the application.
+
+    ── DEEP CHROME, SINCE v5 ───────────────────────────────────────────────────────────────────────────────
+    This rail's own surface moves from Sidebar to DeepChrome in this pass, matching the desktop's NavRail —
+    which is exactly what this control is standing in for on a wide phone, and which took the same v5b move
+    itself: Palette.axaml's own remark on DeepChrome names "the rail down the side" as one of the five things
+    that colour is for. Sidebar stays what a card sits on; this is the frame the cards sit inside, the same
+    distinction PhoneShell's header, strip and bottom bar are drawing at the same time.
+
+    The hover fill moves too, from Panel to Track — and Track rather than Hover, because Palette.axaml's own
+    remark on Track names "the rail row a pointer sits over" as one of the three things that colour is for,
+    literally. Panel was never named for this; it read close enough that nobody had gone looking for the
+    seam, which is exactly the kind of gap Track exists to close.
   -->
 
   <UserControl.Styles>
@@ -54,7 +66,7 @@
     </Style>
 
     <Style Selector="Button.railentry:pointerover /template/ ContentPresenter">
-      <Setter Property="Background" Value="{StaticResource Panel}" />
+      <Setter Property="Background" Value="{StaticResource Track}" />
     </Style>
 
     <Style Selector="Button.railentry.active /template/ ContentPresenter">
@@ -96,7 +108,7 @@
     </Style>
   </UserControl.Styles>
 
-  <Border Width="188" Background="{StaticResource Sidebar}"
+  <Border Width="188" Background="{StaticResource DeepChrome}"
           BorderBrush="{StaticResource Border}" BorderThickness="0,0,1,0">
 
     <!--
@@ -134,15 +146,17 @@
 
           <!--
             The terminal is a surface rather than a page, so it goes through its own command — the same
-            reason the bottom bar's middle entry does. It is never lit, and deliberately: this rail is
-            collapsed while a shell is showing, so IsCurrent could only ever read false. What marks that
-            destination as current is the surface filling the screen.
+            reason the bottom bar's middle entry does. It lights on IsTerminalSurface, which on this control
+            can only mean one thing: the rail stands down for a shell, so the one moment it is drawn beside
+            this surface is Connections with nothing running. That screen is a page in everything but which
+            enum it is in, and a rail sitting beside it with no entry lit would be pointing nowhere.
 
             No count, unlike the two above it. The shells strip sits above the bottom of every page on this
             surface and is that count, drawn as the sessions themselves — a number beside this word would be
             the same fact said twice, three inches apart.
           -->
-          <Button Classes="railentry" Command="{Binding ShowTerminalCommand}">
+          <Button Classes="railentry" Classes.active="{Binding IsTerminalSurface}"
+                  Command="{Binding ShowTerminalCommand}">
             <Grid ColumnDefinitions="Auto,*,Auto">
               <TextBlock Grid.Column="0" Classes="railglyph" Text="⌗" />
               <TextBlock Grid.Column="1" Classes="raillabel" Text="Connections" />

src/DodoSSH.Client.Android/Views/PhoneShell.axaml

@@ -25,8 +25,9 @@
 
     ── a terminal gets the screen ─────────────────────────────────────────────────────────────────────────
     Three of the four rows below stand down while a shell is showing: the header, the shells strip and the
-    bottom bar itself. All three are bound on IsShowingPages, which is the same question asked once — the
-    surface is either a page or a terminal, and these are the chrome a page has.
+    bottom bar itself. The strip asks IsShowingPages directly; the other two go through flags the control
+    computes, and the bar's is the one that differs — it stays up on Connections with nothing running, which
+    is the terminal surface drawing a page rather than a shell. See PhoneShell.RefreshChrome.
 
     The arithmetic is why. Header 56, strip 46, bar 64, and the terminal's own two rows on top of that: at
     360dp the shell was framed by about a third of the display, all of it about somewhere the user was not.
@@ -96,9 +97,18 @@
         the surface. See PhoneShell.ShowsVaultHeader.
       -->
       <Panel Grid.Row="0" IsVisible="{Binding $parent[views:PhoneShell].ShowsVaultHeader}">
-        <Border Background="{StaticResource Chrome}" BorderBrush="{StaticResource Border}"
+        <!--
+          ◆ DeepChrome rather than Chrome, since v5. The desktop's own titlebar and nav rail made the same
+          move in v5b — one step darker than Chrome, #0B0B14 against #10111E, so the frame reads as what
+          holds the glass rather than as another pane of it — and this header is the phone's equivalent
+          furniture: it is what the desktop's titlebar is, on the surface that has no window to carry one.
+          HostActionBar, which takes this header's own place while hosts are selected, and the editor's
+          header in HostsScreen, which the header stands down for, both move with it for the same reason;
+          see the remark on each.
+        -->
+        <Border Background="{StaticResource DeepChrome}" BorderBrush="{StaticResource Border}"
                 BorderThickness="0,0,0,1" Padding="14,0" Height="56">
-          <Grid ColumnDefinitions="Auto,*,Auto,Auto,Auto">
+          <Grid ColumnDefinitions="Auto,*,Auto,Auto">
 
             <!--
               Filled rather than outlined since v2. The mark is the one thing on this header that is not a
@@ -127,30 +137,18 @@
             </StackPanel>
 
             <!--
-              ◆ THE PENCIL, and it is here rather than on the row for the reason the whole gesture split
-              happened: choosing a host on this head no longer raises a bar carrying EDIT, because that bar
-              was five controls over the bottom of the list in the way of a tap that means "connect". EDIT is
-              the one of the five common enough to be worth a control that is always in the same place, so it
-              is in the header — the phone's only piece of persistent chrome — and the long press still
-              reaches the other four. See HostsScreen.axaml.
+              ◆ THE PENCIL WAS HERE, AND IT MOVED INTO THE ACTION BAR.
 
-              Two conditions, nested rather than combined, because Avalonia's bindings have no "and" and the
-              two belong to different view models: which screen is showing is the shell's question, and
-              whether there is a host to edit is the vault's. That is the same arrangement the header itself
-              is wrapped in one level up.
+              It was in the header because a long press raised a connect card carrying EDIT, and a card over
+              the bottom of the list was the wrong place for the one action people reach for most. Both
+              halves of that changed at once: a long press now chooses hosts rather than raising a card, and
+              the bar it raises takes this header's place — so the pencil is in that bar, beside the count of
+              what it would edit. See HostActionBar.axaml, which is drawn in the Panel below this one.
 
-              Collapsed rather than disabled when there is nothing chosen. A greyed pencil sitting beside the
-              vault's name on every screen would be a permanent reminder of a control that is only ever about
-              one row.
+              What is left in this row is the vault's name, the sync light and LOCK: three facts about the
+              keychain, and none of them about a row.
             -->
-            <Panel Grid.Column="3" IsVisible="{Binding IsHostsShowing}">
-              <Button Classes="icon" Content="✎" FontSize="15" Margin="4,0,0,0"
-                      IsVisible="{Binding Vault.CanEditSelectedHost}"
-                      Command="{Binding Vault.EditSelectedHostCommand}"
-                      ToolTip.Tip="Edit the selected host" />
-            </Panel>
-
-            <Button Grid.Column="4" Classes="icon" Margin="4,0,0,0" Command="{Binding LockCommand}"
+            <Button Grid.Column="3" Classes="icon" Margin="4,0,0,0" Command="{Binding LockCommand}"
                     ToolTip.Tip="Lock the keychain">
               <TextBlock Text="LOCK" Classes="label" FontSize="8.5"
                          Foreground="{StaticResource TextDim}" />
@@ -159,6 +157,21 @@
         </Border>
       </Panel>
 
+      <!-- ============ ◆ the contextual action bar ============ -->
+      <!--
+        In the header's row and in its place, never beside it. A bar that stacked under the vault's name
+        would be two rows of chrome above a list at 360dp, and worse, it would leave the keychain's name and
+        LOCK on screen over a list that is no longer about the keychain — see PhoneShell.ShowsHostSelectionBar,
+        which is where the two are made exclusive.
+
+        Wrapped so the data context can be the vault, as every other screen in this file is: what the bar
+        counts and what its four controls do are one vault's business, and the shell's own binding would
+        resolve none of them.
+      -->
+      <Panel Grid.Row="0" IsVisible="{Binding $parent[views:PhoneShell].ShowsHostSelectionBar}">
+        <views:HostActionBar DataContext="{Binding Vault}" />
+      </Panel>
+
       <!-- The screens. Only one draws; which one is the shell's business. -->
       <Panel Grid.Row="1">
         <!--
@@ -252,6 +265,12 @@
         ◆ The two host-key decisions, over everything. Outside the screen Panel deliberately: file transfer
         makes its own trust decision on the desktop, and the phone now has that screen — so a sheet nested
         inside the hosts screen would be one the files screen could not raise.
+
+        ◆ And that arrangement is what the shell now relies on rather than merely permits. It used to move to
+        HOSTS before letting a handshake ask the question, because the desktop head drew it as a banner on that
+        screen; it does not any more, so this sheet is what is over the Connections surface when a machine
+        typed into the connect box by hand turns out to be a first contact. Which is the case the navigation
+        was worst for — it took the box away. See MainWindowViewModel.OnVaultConnectionFailed.
       -->
       <Panel Grid.Row="1">
         <views:HostKeySheet DataContext="{Binding Vault}" />
@@ -271,9 +290,17 @@
         two rows of the same pills — one of them 46 pixels of it — is the arrangement this surface exists to
         stop. Wrapped rather than given a second condition, because the strip's own visibility is about
         whether there are any tabs and this one is about which surface is up.
+
+        ◆ And it stands down under the host editor, which is a page rather than a card since the pencil moved
+        into the action bar. A strip of open shells above a form is the same two-rows-of-chrome problem the
+        terminal has, on a screen where the form is the whole point of being there. See
+        PhoneShell.ShowsShellStrip, which is where that "and" is made, Avalonia's bindings having none.
       -->
-      <Panel Grid.Row="2" IsVisible="{Binding IsShowingPages}">
-        <Border IsVisible="{Binding HasTabs}" Background="{StaticResource Sidebar}"
+      <Panel Grid.Row="2" IsVisible="{Binding $parent[views:PhoneShell].ShowsShellStrip}">
+        <!-- DeepChrome rather than Sidebar, since v5, joining the header and the bar above and below it:
+             the strip is chrome the same way they are — a frame around the screen rather than a pane of
+             it — and Sidebar is what a card sits on, which this row is not. -->
+        <Border IsVisible="{Binding HasTabs}" Background="{StaticResource DeepChrome}"
                 BorderBrush="{StaticResource Border}" BorderThickness="0,1,0,0" Height="46">
           <ScrollViewer HorizontalScrollBarVisibility="Auto" VerticalScrollBarVisibility="Disabled">
             <ItemsControl ItemsSource="{Binding Tabs}" Margin="12,0" VerticalAlignment="Center">
@@ -282,7 +309,7 @@
               </ItemsControl.ItemsPanel>
               <ItemsControl.ItemTemplate>
                 <DataTemplate x:DataType="vm:TerminalTabViewModel">
-                  <Button Classes="row" MinHeight="34" Padding="13,0" CornerRadius="9"
+                  <Button Classes="row" MinHeight="34" Padding="13,0" CornerRadius="10"
                           Background="{StaticResource Panel}" BorderBrush="{StaticResource BorderMid}"
                           BorderThickness="1"
                           Command="{Binding $parent[views:PhoneShell].((vm:MainWindowViewModel)DataContext).SelectTabCommand}"
@@ -293,8 +320,14 @@
                         which was true when a tab could not exist without a session; one can now —
                         connecting opens the tab first — and a dot that was green before anything had
                         answered would be the one thing on this strip claiming something untrue.
+
+                        Amber while it is being made, which is the other half of that correction. Not being
+                        green stopped the dot lying, but it left a tab still dialling drawn exactly like a
+                        tab whose shell has exited — the two states on this strip with the least in common,
+                        one worth waiting for and one over. See Phone.axaml.
                       -->
-                      <Ellipse Classes="dot" Classes.live="{Binding IsLive}" Width="6" Height="6"
+                      <Ellipse Classes="dot" Classes.live="{Binding IsLive}"
+                               Classes.connecting="{Binding IsConnecting}" Width="6" Height="6"
                                VerticalAlignment="Center" />
                       <TextBlock Classes="mono" FontSize="11" Text="{Binding Label}" />
                     </StackPanel>
@@ -308,10 +341,15 @@
 
       <!-- ============ navigation ============ -->
       <!--
-        Gone while a terminal is showing, which is the whole of that surface's arrangement: the bar's three
+        Gone while a shell is showing, which is the whole of that surface's arrangement: the bar's three
         destinations are replaced by a back arrow and a + that leads to all of them, both in the terminal's
         own bar. See TerminalScreen.axaml.
 
+        ◆ A shell rather than the surface, which are not the same question — Connections with nothing running
+        is that surface drawing a page, and the bar stays under it. The screen it stands down for is the one
+        that pays for the room; a box asking which machine to connect to does not, and taking the nav away
+        from somebody who has just closed their last tab left them with only the back gesture.
+
         ◆ It is no longer a single question — a wide surface takes the rail instead — so it reads one flag
         the control computes rather than a condition here. See PhoneShell.ShowsBottomBar.
 
@@ -321,8 +359,10 @@
         shape of everything else already behind the hub. What is left is the two halves of using this
         application, and the drawer holding the rest.
       -->
+      <!-- DeepChrome rather than Chrome, since v5 — see the remark on the vault header, above, which this
+           bar is the foot of the same frame the header is the top of. -->
       <Border Grid.Row="3" IsVisible="{Binding $parent[views:PhoneShell].ShowsBottomBar}"
-              Background="{StaticResource Chrome}" BorderBrush="{StaticResource Border}"
+              Background="{StaticResource DeepChrome}" BorderBrush="{StaticResource Border}"
               BorderThickness="0,1,0,0" Height="64">
         <Grid ColumnDefinitions="*,*,*">
 
@@ -340,12 +380,15 @@
             rather than the door. The enum member stays ShellSurface.Terminal, for the reason the keychain's
             tab is not called Vault: the surface is a terminal, and the word a user reads is the product's.
 
-            The only entry here that never lights, and deliberately no longer tries: this bar is collapsed
-            while the terminal is showing, so IsCurrent could only ever be read as false. Binding it anyway
-            would be a rule about a state this control cannot be in. What marks it as current is the surface
-            filling the screen.
+            ◆ It lights again, and IsTerminalSurface is the right question rather than a near miss. This bar
+            is drawn in exactly two situations — a page, and Connections with nothing running — and that flag
+            separates them: false for every page, true for the one screen where this is where you are. With a
+            shell up the bar is gone and nothing here is read at all. It was left unbound while the bar was
+            collapsed for the whole of the terminal surface, which made a lit state unreachable; a screen the
+            bar can now sit under and not point at is the entry looking broken instead.
           -->
           <views:NavButton Grid.Column="1" Label="Connections" Glyph="⌗"
+                           IsCurrent="{Binding IsTerminalSurface}"
                            Command="{Binding ShowTerminalCommand}" />
 
           <!--

src/DodoSSH.Client.Android/Views/PhoneShell.axaml.cs

@@ -17,6 +17,18 @@ internal sealed partial class PhoneShell : UserControl
 {
     private MainWindowViewModel? shell;
 
+    /// <summary>
+    /// The open vault, while there is one, so that this control hears about the hosts screen's own state.
+    /// </summary>
+    /// <remarks>
+    /// ◆ <b>A second subscription, and it is the price of the header being swappable.</b> Two of the flags
+    /// below are questions about the vault rather than about the shell — whether hosts are ticked, and
+    /// whether the host editor is filling the screen — and the shell does not forward the vault's
+    /// notifications. Kept in step from <see cref="OnShellChanged"/>, because <c>Vault</c> is replaced on
+    /// every unlock and nulled on every lock; a handler left on a disposed vault would keep it alive.
+    /// </remarks>
+    private VaultViewModel? vault;
+
     /// <summary>
     /// Everything the phone draws, which is the element the software keyboard is kept off.
     /// </summary>
@@ -90,10 +102,40 @@ internal sealed partial class PhoneShell : UserControl
                 TryOfferDeviceUnlock();
             }
 
+            FollowTheVault();
             RefreshChrome();
         };
     }
 
+    /// <summary>Moves this control's second subscription onto whichever vault is open now.</summary>
+    /// <remarks>
+    /// Compared before being swapped, so that the ordinary case — a shell notification about something else
+    /// entirely — costs one reference comparison rather than an unsubscribe and a resubscribe per property
+    /// change on the shell.
+    /// </remarks>
+    private void FollowTheVault()
+    {
+        if (ReferenceEquals(vault, shell?.Vault))
+        {
+            return;
+        }
+
+        if (vault is not null)
+        {
+            vault.PropertyChanged -= OnVaultChanged;
+        }
+
+        vault = shell?.Vault;
+
+        if (vault is not null)
+        {
+            vault.PropertyChanged += OnVaultChanged;
+        }
+    }
+
+    private void OnVaultChanged(object? sender, System.ComponentModel.PropertyChangedEventArgs e)
+        => RefreshChrome();
+
     /// <summary>Whether this surface is wide enough to be laid out like the desktop.</summary>
     /// <remarks>
     /// A property of the control rather than of the view model, because it is a fact about the surface and
@@ -104,10 +146,19 @@ internal sealed partial class PhoneShell : UserControl
         AvaloniaProperty.Register<PhoneShell, bool>(nameof(IsWide));
 
     /// <summary>Whether the rail down the left edge is drawn.</summary>
+    /// <remarks>
+    /// The wide surface's answer to <see cref="ShowsBottomBarProperty"/> and asks the same question about
+    /// which screen is up, so the two move together — including over Connections with nothing running. See
+    /// <see cref="RefreshChrome"/>.
+    /// </remarks>
     public static readonly StyledProperty<bool> ShowsRailProperty =
         AvaloniaProperty.Register<PhoneShell, bool>(nameof(ShowsRail));
 
     /// <summary>Whether the three-entry bar across the bottom is drawn.</summary>
+    /// <remarks>
+    /// Not simply the pages: it also stays up on Connections with nothing running, which is the terminal
+    /// surface drawing a page. See <see cref="RefreshChrome"/>.
+    /// </remarks>
     public static readonly StyledProperty<bool> ShowsBottomBarProperty =
         AvaloniaProperty.Register<PhoneShell, bool>(nameof(ShowsBottomBar));
 
@@ -115,6 +166,14 @@ internal sealed partial class PhoneShell : UserControl
     public static readonly StyledProperty<bool> ShowsVaultHeaderProperty =
         AvaloniaProperty.Register<PhoneShell, bool>(nameof(ShowsVaultHeader));
 
+    /// <summary>Whether the bar about the chosen hosts is drawn in the header's place.</summary>
+    public static readonly StyledProperty<bool> ShowsHostSelectionBarProperty =
+        AvaloniaProperty.Register<PhoneShell, bool>(nameof(ShowsHostSelectionBar));
+
+    /// <summary>Whether the strip of open shells above the bottom bar is drawn.</summary>
+    public static readonly StyledProperty<bool> ShowsShellStripProperty =
+        AvaloniaProperty.Register<PhoneShell, bool>(nameof(ShowsShellStrip));
+
     /// <inheritdoc cref="IsWideProperty" />
     public bool IsWide
     {
@@ -143,17 +202,40 @@ internal sealed partial class PhoneShell : UserControl
         private set => SetValue(ShowsVaultHeaderProperty, value);
     }
 
+    /// <inheritdoc cref="ShowsHostSelectionBarProperty" />
+    public bool ShowsHostSelectionBar
+    {
+        get => GetValue(ShowsHostSelectionBarProperty);
+        private set => SetValue(ShowsHostSelectionBarProperty, value);
+    }
+
+    /// <inheritdoc cref="ShowsShellStripProperty" />
+    public bool ShowsShellStrip
+    {
+        get => GetValue(ShowsShellStripProperty);
+        private set => SetValue(ShowsShellStripProperty, value);
+    }
+
     /// <summary>
     /// Works out which chrome this surface should be wearing.
     /// </summary>
     /// <remarks>
     /// <para>
-    /// ◆ <b>Three flags computed here rather than three conditions in the markup, because Avalonia's
+    /// ◆ <b>Five flags computed here rather than five conditions in the markup, because Avalonia's
     /// bindings have no "and" and none of these is a single question any more.</b> Everywhere else on this
     /// head that costs a wrapper element; here it would cost two nested ones per row and the header's would
     /// have to be an "or", which a wrapper cannot express at all.
     /// </para>
     /// <para>
+    /// ◆ <b>The nav stands down for a shell rather than for the terminal surface, and those parted company
+    /// when that surface gained a page.</b> Connections with nothing running is a box, a CONNECT button and
+    /// the machines connected to before — see TerminalScreen.axaml — and none of that is worth the screen a
+    /// shell is worth it for. It is also the one screen somebody can arrive at by closing their last tab,
+    /// which made the collapsed bar a way to end up with no route to Hosts or Settings but the system back
+    /// gesture. The header is deliberately not part of this: the surface draws its own bar with the back
+    /// arrow and the +, and the vault header above that is the second row this head exists to avoid.
+    /// </para>
+    /// <para>
     /// <b>The header is the one worth reading twice.</b> Narrow, it stands down behind SETTINGS, because the
     /// screens under that hub draw their own header with a back arrow and two rows of chrome is what this
     /// surface exists to avoid. Wide, there is no hub to be behind and no back arrow to duplicate — the rail
@@ -161,7 +243,16 @@ internal sealed partial class PhoneShell : UserControl
     /// Losing them on the keychain would be losing the only LOCK button on the surface.
     /// </para>
     /// <para>
-    /// Recomputed on every shell notification rather than on a named list of them. Three boolean
+    /// ◆ <b>The header is now a swap rather than a switch, and the editor takes the whole screen.</b> Two
+    /// more flags and two more inputs, both of them the vault's rather than the shell's — see
+    /// <see cref="vault"/>. While hosts are ticked the header stands down and
+    /// <see cref="ShowsHostSelectionBar"/> puts the action bar in its place, which is what makes that bar
+    /// unambiguous: the screen is about the ticked hosts and nothing else. While the host editor is open it
+    /// is a page rather than a card, so all four rows of chrome stand down and the form has the display —
+    /// which is what "opens in a separate page" means on a 360dp screen.
+    /// </para>
+    /// <para>
+    /// Recomputed on every notification from either object rather than on a named list of them. Five boolean
     /// comparisons and no allocation is cheaper than being wrong: the properties this reads are computed
     /// ones, and which of them raise a change is a fact about a file in another project that nothing here
     /// would notice going stale.
@@ -172,6 +263,24 @@ internal sealed partial class PhoneShell : UserControl
         var wide = body.Bounds.Width >= WideAt;
         var pages = shell?.IsShowingPages == true;
 
+        // ◆ Connections with nothing running, which is the terminal surface drawing a page: a box, a CONNECT
+        // button and the machines connected to before. The nav stands down for a shell — the whole of the
+        // arrangement below — and there is no shell here to stand down for, so it stays. Taking it away on
+        // this one screen was worst where it was least affordable: somebody who has just closed their last
+        // tab, or who pressed Connections to see what was open and found nothing, was left on a screen whose
+        // only way to Hosts or Settings was the system back gesture.
+        var connectPage = shell is { IsTerminalSurface: true, HasTabs: false };
+
+        // The editor is a page of its own now, so nothing else is drawn around it — not the vault header,
+        // not the shells strip, and not the way off the screen. Its own header carries the back arrow, which
+        // is the one control it needs and the one the system gesture already maps to.
+        var editing = vault?.IsEditing == true;
+
+        // Only on the hosts screen. The ticks survive a trip to the keychain — the set is not cleared by
+        // navigating — and a bar counting hosts over the transfers screen would be chrome about a list that
+        // is not on the display.
+        var choosing = vault?.IsChoosingHosts == true && shell?.IsHostsShowing == true;
+
         // Before the flags, because it changes what one of them reads. Nothing else on this head navigates
         // in response to a resize, and this is not navigation for its own sake: the hub is a list of the
         // destinations the rail now carries, so an unfolded device would otherwise sit on a menu of things
@@ -183,9 +292,11 @@ internal sealed partial class PhoneShell : UserControl
         }
 
         IsWide = wide;
-        ShowsRail = wide && pages;
-        ShowsBottomBar = !wide && pages;
-        ShowsVaultHeader = pages && (wide || shell?.IsMoreSurface != true);
+        ShowsRail = wide && (pages || connectPage) && !editing;
+        ShowsBottomBar = !wide && (pages || connectPage) && !editing;
+        ShowsShellStrip = pages && !editing;
+        ShowsHostSelectionBar = pages && choosing && !editing;
+        ShowsVaultHeader = pages && !editing && !choosing && (wide || shell?.IsMoreSurface != true);
     }
 
     private void OnShellChanged(object? sender, System.ComponentModel.PropertyChangedEventArgs e)
@@ -195,6 +306,7 @@ internal sealed partial class PhoneShell : UserControl
             return;
         }
 
+        FollowTheVault();
         RefreshChrome();
 
         if (e.PropertyName is nameof(MainWindowViewModel.State))
@@ -279,11 +391,15 @@ internal sealed partial class PhoneShell : UserControl
                 ApplyKeyboardInset(keyboard);
             }
         }
+
+        PhoneEnvironment.Resumed += OnActivityResumed;
     }
 
     /// <inheritdoc />
     protected override void OnDetachedFromVisualTree(VisualTreeAttachmentEventArgs e)
     {
+        PhoneEnvironment.Resumed -= OnActivityResumed;
+
         if (TopLevel.GetTopLevel(this) is { } top)
         {
             top.BackRequested -= OnBackRequested;
@@ -298,6 +414,43 @@ internal sealed partial class PhoneShell : UserControl
         base.OnDetachedFromVisualTree(e);
     }
 
+    /// <summary>
+    /// Re-reads the keyboard inset on the way back into the application.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// ◆ <b>The margin can be stale here and there is no other moment it gets corrected.</b> Android takes
+    /// the software keyboard away when this activity stops, and Avalonia's input pane raises no
+    /// <c>StateChanged</c> for it — the pane is gone with the window rather than closed in front of one. So
+    /// <see cref="ApplyKeyboardInset(double)"/> last ran for a keyboard that was up, and the interface comes
+    /// back holding a whole keyboard's worth of empty space it no longer needs. It is not subtle and it does
+    /// not recover: a third of the display, on every screen, until something opens and closes the keyboard.
+    /// </para>
+    /// <para>
+    /// Asking the pane rather than assuming zero, because resuming does not always mean the keyboard went.
+    /// A permission dialogue, the biometric sheet and the sign-in browser all pause this activity and can
+    /// all hand it back with the keyboard exactly as it was — and a handler that flattened the inset on the
+    /// way in would then strand the focused box under a keyboard that is genuinely there. The pane knows
+    /// which of the two happened; this only has to ask it again.
+    /// </para>
+    /// <para>
+    /// Posted rather than run inline. <c>OnResume</c> is early — the window has not necessarily been laid
+    /// out again and the pane's own state is settled by the insets that arrive with the next frame, so
+    /// reading it on this turn can return the state that is on the way out. Background priority puts the
+    /// read after that.
+    /// </para>
+    /// </remarks>
+    private void OnActivityResumed(object? sender, EventArgs e) =>
+        Dispatcher.UIThread.Post(
+            () =>
+            {
+                if (keyboard is not null)
+                {
+                    ApplyKeyboardInset(keyboard);
+                }
+            },
+            DispatcherPriority.Background);
+
     private void OnKeyboardChanged(object? sender, InputPaneStateEventArgs e)
         => ApplyKeyboardInset(e.NewState is InputPaneState.Open ? e.EndRect.Height : 0);
 
@@ -500,16 +653,23 @@ internal sealed partial class PhoneShell : UserControl
     /// <returns>Whether anything was closed, and so whether back has been spent.</returns>
     /// <remarks>
     /// <para>
-    /// Order is the whole of it. The two sheets sit over the list and the two editors sit in place of it, so
-    /// a sheet has to go first — closing an editor while a sheet was open would leave the sheet floating
-    /// over a list nobody asked to see, and the second back would then close the sheet rather than the
-    /// editor the user was looking at.
+    /// Order is the whole of it. The sheets sit over the list, the panels sit above it and the editors sit
+    /// in place of it, so a sheet has to go first — closing an editor while a sheet was open would leave the
+    /// sheet floating over a list nobody asked to see, and the second back would then close the sheet rather
+    /// than the editor the user was looking at.
     /// </para>
     /// <para>
     /// The editors are cancelled rather than merely hidden. Cancelling is what clears the boxes, and the
     /// host editor's boxes are the ones worth clearing: leaving a half-typed hostname behind would have the
     /// next NEW HOST open on somebody else's abandoned draft.
     /// </para>
+    /// <para>
+    /// ◆ <b>Selection mode is last and is still a thing back has to spend itself on.</b> It is a mode rather
+    /// than a surface — the list underneath is fully drawn and the only sign of it is the bar across the top
+    /// — and a gesture that left the application from it would take somebody out of the app because they had
+    /// held a row down. Its panels go before it, in the order they are stacked: the picker or the question is
+    /// what the user is looking at, and the ticks underneath are what it is about.
+    /// </para>
     /// </remarks>
     private static bool TryCloseAnOpenEditor(MainWindowViewModel current)
     {
@@ -518,22 +678,71 @@ internal sealed partial class PhoneShell : UserControl
             return false;
         }
 
+        return TryLowerASheet(vault) || TryCloseSomethingBehindTheSheets(vault);
+    }
+
+    /// <summary>Lowers the nearest of the four sheets, which are what sits over everything else.</summary>
+    /// <remarks>
+    /// The four cannot be open at once — each is raised from a control the others hide — so their order
+    /// between themselves decides nothing. What matters is that all of them come before the panels and the
+    /// editors underneath: closing an editor while a sheet was open would leave the sheet floating over a
+    /// list nobody asked to see.
+    /// </remarks>
+    private static bool TryLowerASheet(VaultViewModel vault)
+    {
+        // ◆ The action bar's own menu, first of the four because it is raised from chrome that is already
+        // over everything else.
+        if (vault.IsHostActionSheetOpen)
+        {
+            vault.CloseHostActionSheetCommand.Execute(null);
+            return true;
+        }
+
+        // The password sheet, which is what a tap on a machine that wants one raises. Cancelled rather than
+        // hidden, because cancelling is what empties the box — see VaultViewModel.CancelConnectPassword.
+        if (vault.IsAskingForConnectPassword)
+        {
+            vault.CancelConnectPasswordCommand.Execute(null);
+            return true;
+        }
+
         if (vault.IsAddSheetOpen)
         {
             vault.CloseAddSheetCommand.Execute(null);
             return true;
         }
 
-        // The other sheet, and it is checked beside the first rather than after the editors for the same
-        // reason: it is raised over the list, so it is the nearest thing on screen. The two cannot be open
-        // at once — one is raised by the +, the other by a heading, and each hides the list the other's
-        // control is on — so their order between themselves decides nothing.
         if (vault.GroupSheet is not null)
         {
             vault.CloseGroupSheetCommand.Execute(null);
             return true;
         }
 
+        return false;
+    }
+
+    /// <summary>Closes the nearest of the panels, the editors and selection mode itself.</summary>
+    /// <inheritdoc cref="TryCloseAnOpenEditor" path="/remarks" />
+    private static bool TryCloseSomethingBehindTheSheets(VaultViewModel vault)
+    {
+        if (vault.IsSendingChosenHostsToAVault)
+        {
+            vault.CancelSendChosenHostsToAVaultCommand.Execute(null);
+            return true;
+        }
+
+        if (vault.IsRegroupingChosenHosts)
+        {
+            vault.CancelRegroupChosenHostsCommand.Execute(null);
+            return true;
+        }
+
+        if (vault.IsConfirmingChosenHostDeletion)
+        {
+            vault.CancelDeleteCommand.Execute(null);
+            return true;
+        }
+
         if (vault.IsEditing)
         {
             vault.CancelEditCommand.Execute(null);
@@ -546,6 +755,12 @@ internal sealed partial class PhoneShell : UserControl
             return true;
         }
 
+        if (vault.IsChoosingHosts)
+        {
+            vault.ClearHostChoiceCommand.Execute(null);
+            return true;
+        }
+
         return false;
     }
 

src/DodoSSH.Client.Android/Views/PreferencesScreen.axaml

@@ -163,6 +163,8 @@
                      Text="Text size is on the terminal itself, where the text is — the phone puts A− and A+ in the shell's own bar rather than a step away from what they resize." />
           <TextBlock Classes="body"
                      Text="Importing an ~/.ssh/config has no meaning on Android: scoped storage means there is no such file for this application to read." />
+          <TextBlock Classes="body"
+                     Text="Importing a Hetzner Cloud project would work here — it is a token and a network request, not a file — and the wizard behind it is already shared. What is missing is a phone layout for it: the desktop draws a five-column table, and the screen it would become is not the desktop's with narrower columns. Until then it is a thing to do on a desktop; the hosts it makes sync here like any other." />
           <TextBlock Classes="body"
                      Text="Signing out is on the lock screen, where somebody who has forgotten their passphrase can reach it without unlocking anything first." />
         </StackPanel>

src/DodoSSH.Client.Android/Views/RecoveryCodeScreen.axaml

@@ -36,8 +36,16 @@
         good way to read a code and a bad way to copy one: a user who cannot select it will photograph the
         screen, and that is a worse home for it than their clipboard.
       -->
+      <!--
+        ◆ CornerRadius 12, not 6. Its padding, 14,13, is Border.card's own — this box was already drawn to a
+        card's proportions, in every dimension but the corner — so under v2's ladder 6 was an outlier this
+        file never explained. It is worse than unexplained now: v5's ladder gives 6 to a chip, and a
+        full-width box holding the one secret this whole screen exists to show is not a chip wearing a
+        card's padding by coincidence. 12 is what a card-shaped box gets on this ladder, and this one always
+        was one.
+      -->
       <Border Margin="0,20,0,0" Background="{StaticResource Field}" BorderBrush="{StaticResource BorderMid}"
-              BorderThickness="1" CornerRadius="6" Padding="14,13">
+              BorderThickness="1" CornerRadius="12" Padding="14,13">
         <SelectableTextBlock Text="{Binding RecoveryCode}"
                              FontFamily="{StaticResource MonoFont}" FontSize="14"
                              Foreground="{StaticResource Accent}"

src/DodoSSH.Client.Android/Views/ServerScreen.axaml

@@ -20,9 +20,17 @@
       <StackPanel Spacing="10" HorizontalAlignment="Center" Margin="0,48,0,36">
         <!--
           Filled rather than outlined since v2, the same mark the header, the desktop titlebar and the
-          launcher icon carry. The radius is not picked: the mark runs 6 at 20 and 8 at 26, which is a
-          third of a unit per unit of tile and lands on 14 at 44 — and 14 is a rung of Phone.axaml's
-          ladder, the one for a block of monospaced output, which is what this is.
+          launcher icon carry. The radius is not picked: the mark runs 6 at 20 and 8 at 26, a third of a
+          unit per unit of tile, and lands on 14 at 44.
+
+          ◆ That used to be a rung of Phone.axaml's own ladder too — 14, the one for a block of monospaced
+          output, which this glyph reads as — and the coincidence was worth a sentence: one fewer number
+          the ladder had to introduce. It is a coincidence no longer. v5's ladder moves output to 12 and
+          never reaches 14 at all, so the mark's own 6/8/14 progression stands on its own now rather than
+          borrowing a card- or output-shaped justification — a self-contained proportion for a badge that
+          answers to nothing but its own three sizes. Left at 14 rather than moved to 12 with everything
+          else: the ladder governs content surfaces, and this is a mark, sized off its own tile rather than
+          off what it holds.
         -->
         <Border Width="44" Height="44" CornerRadius="14" Background="{StaticResource Accent}"
                 HorizontalAlignment="Center">

src/DodoSSH.Client.Android/Views/SnippetsScreen.axaml

@@ -64,6 +64,26 @@
 
           <TextBox Classes="field" Text="{Binding EditorLabel}" PlaceholderText="name" />
 
+          <!--
+            ◆ Which vault a *new* snippet is filed into. Hidden for an existing one — its vault is not a
+            field of this form, and changing it is MOVE below — and hidden entirely where there is only one
+            vault to choose between, which is where most people stay.
+          -->
+          <StackPanel Spacing="6" IsVisible="{Binding ShowsEditorVaultChoice}">
+            <TextBlock Classes="label" Text="VAULT" />
+            <ComboBox HorizontalAlignment="Stretch" MinHeight="44"
+                      ItemsSource="{Binding EditorVaultChoices}"
+                      SelectedItem="{Binding EditorSelectedVault}">
+              <ComboBox.ItemTemplate>
+                <DataTemplate x:DataType="vm:VaultChoiceViewModel">
+                  <TextBlock Classes="mono" FontSize="12" Text="{Binding Display}" />
+                </DataTemplate>
+              </ComboBox.ItemTemplate>
+            </ComboBox>
+            <TextBlock Classes="body"
+                       Text="A shared vault means everybody holding its key can read this command and insert it into their own terminals." />
+          </StackPanel>
+
           <TextBox Classes="field" Text="{Binding EditorCommand}" PlaceholderText="command"
                    Height="120" AcceptsReturn="True" TextWrapping="Wrap"
                    VerticalContentAlignment="Top" Padding="14,10" />
@@ -125,10 +145,23 @@
                           IsVisible="{Binding Badge, Converter={x:Static StringConverters.IsNotNullOrEmpty}}">
                     <TextBlock Text="{Binding Badge}" />
                   </Border>
+
+                  <!--
+                    ◆ Which vault this one is in, drawn only where there is more than one to be in. A shared
+                    snippet is a command the rest of a team can read and insert into their own terminals,
+                    and this card is the only place that fact appears before somebody taps EDIT.
+                  -->
+                  <Border Classes="tag outline" IsVisible="{Binding HasVaultBadge}">
+                    <TextBlock Text="{Binding VaultBadge}" />
+                  </Border>
                 </StackPanel>
 
-                <!-- The command, on the surface every block of monospace in this design is drawn on. -->
-                <Border Classes="output" CornerRadius="9" Padding="11,9">
+                <!-- The command, on the surface every block of monospace in this design is drawn on.
+                     No radius override any more: it diverged from Border.output's own 14 under v2 for a
+                     smaller card-nested block, and now that the style's own value has moved to 12 there is
+                     no gap left worth a second number for — it takes the class's, like every other output
+                     block on this head. -->
+                <Border Classes="output" Padding="11,9">
                   <TextBlock Classes="mono" FontSize="11" Text="{Binding Preview}"
                              Foreground="{StaticResource Text}" TextTrimming="CharacterEllipsis" />
                 </Border>
@@ -145,26 +178,76 @@
       second column to put it in. Both buttons name the terminal — "TYPE INTO pg-primary" — because on a
       phone the tab strip may be scrolled away and "whatever is in the terminal receives this" is not a
       sentence anybody should have to guess the subject of.
+
+      DeepChrome rather than Chrome, since v5: this bar is chrome furniture in the same sense the bottom
+      nav bar it sits above is — a raised strip of frame rather than a pane of content — and joins the same
+      DeepChrome decision PhoneShell's header, strip and bar made. See the remark there.
     -->
-    <Border Grid.Row="4" IsVisible="{Binding HasSelection}" Background="{StaticResource Chrome}"
+    <Border Grid.Row="4" IsVisible="{Binding HasSelection}" Background="{StaticResource DeepChrome}"
             BorderBrush="{StaticResource Border}" BorderThickness="0,1,0,0" Padding="14,12"
             IsEnabled="{Binding !IsEditing}">
       <StackPanel Spacing="9">
 
-        <Button Classes="primary" Content="{Binding InsertLabel}" Command="{Binding InsertCommand}"
-                IsEnabled="{Binding CanInsert}" />
+        <StackPanel Spacing="9" IsVisible="{Binding ShowsSelectionActions}">
+          <Button Classes="primary" Content="{Binding InsertLabel}" Command="{Binding InsertCommand}"
+                  IsEnabled="{Binding CanInsert}" />
 
-        <!-- Only for a snippet marked as running. For every other one there is nothing this could mean. -->
-        <Button Classes="secondary" Height="44" Content="{Binding RunLabel}"
-                Command="{Binding RunCommand}" IsVisible="{Binding SelectionRuns}"
-                IsEnabled="{Binding CanInsert}" />
+          <!-- Only for a snippet marked as running. For every other one there is nothing this could mean. -->
+          <Button Classes="secondary" Height="44" Content="{Binding RunLabel}"
+                  Command="{Binding RunCommand}" IsVisible="{Binding SelectionRuns}"
+                  IsEnabled="{Binding CanInsert}" />
 
-        <Grid ColumnDefinitions="*,8,*">
-          <Button Grid.Column="0" Classes="secondary" Height="44" Content="EDIT"
-                  Command="{Binding EditCommand}" />
-          <Button Grid.Column="2" Classes="danger" Height="44" Content="DELETE"
-                  Command="{Binding DeleteCommand}" />
-        </Grid>
+          <Grid ColumnDefinitions="*,8,*">
+            <Button Grid.Column="0" Classes="secondary" Height="44" Content="EDIT"
+                    Command="{Binding EditCommand}" />
+            <!--
+              ◆ Sharing the snippet. Beside EDIT rather than inside it, which is the line the hosts screen
+              draws too: the two vaults are encrypted under different keys, so this is a re-seal into one and
+              a tombstone in the other — nothing a SAVE could do. It shows only where there is somewhere to
+              move to; see SnippetsViewModel.CanMove.
+            -->
+            <Button Grid.Column="2" Classes="secondary" Height="44" Content="MOVE"
+                    IsVisible="{Binding CanMove}" Command="{Binding MoveCommand}" />
+          </Grid>
+
+          <!--
+            ◆ A row of its own, under the pair rather than beside them, as the hosts screen puts it. A phone
+            has no hover and no tooltip, so the only thing separating a destructive button from an ordinary
+            one is where a thumb lands.
+          -->
+          <Button Classes="danger" Height="44" Content="DELETE" Command="{Binding DeleteCommand}" />
+        </StackPanel>
+
+        <!--
+          ◆ MOVING THE SNIPPET TO ANOTHER VAULT, in the place the insert controls were. A picker and two
+          buttons rather than a question with a yes: what is being asked is which vault, and a move is undone
+          by moving it back.
+
+          The sentence is not decoration. Unlike a host, a snippet crosses whole — there is no group and no
+          tag to leave behind — so what there is to say is who can read it afterwards; and on a phone, where
+          the status line is one line at the bottom of a screen somebody has already navigated away from,
+          before the tap is the only place that reliably gets read.
+        -->
+        <StackPanel Spacing="10" IsVisible="{Binding IsMoving}">
+          <TextBlock Classes="label" Text="MOVE TO VAULT" />
+          <ComboBox HorizontalAlignment="Stretch" MinHeight="44"
+                    ItemsSource="{Binding MoveVaultChoices}"
+                    SelectedItem="{Binding SelectedMoveVault}">
+            <ComboBox.ItemTemplate>
+              <DataTemplate x:DataType="vm:VaultChoiceViewModel">
+                <TextBlock Classes="mono" FontSize="12" Text="{Binding Display}" />
+              </DataTemplate>
+            </ComboBox.ItemTemplate>
+          </ComboBox>
+          <TextBlock Classes="body"
+                     Text="The snippet is re-encrypted with the other vault's key, so everybody who holds that key can read this command and insert it — and nobody else can. Nothing else about it changes." />
+          <Grid ColumnDefinitions="*,8,*">
+            <Button Grid.Column="0" Classes="primary" Height="44" Content="MOVE"
+                    Command="{Binding ConfirmMoveCommand}" />
+            <Button Grid.Column="2" Classes="secondary" Height="44" Content="CANCEL"
+                    Command="{Binding CancelMoveCommand}" />
+          </Grid>
+        </StackPanel>
       </StackPanel>
     </Border>
 

src/DodoSSH.Client.Android/Views/TerminalInputEventArgs.cs

@@ -0,0 +1,16 @@
+namespace DodoSSH.Client.Android.Views;
+
+/// <summary>
+/// One keystroke's worth of bytes, on its way from <see cref="TerminalInputSurface"/> to the shell.
+/// </summary>
+/// <remarks>
+/// Bytes rather than text, because half of what a terminal receives has no text form: a cursor key is an
+/// escape sequence and Ctrl-C is a single byte below the printable range. Encoding at the surface — where
+/// the key is still known — rather than downstream is what keeps the send path from having to guess which
+/// kind it was handed.
+/// </remarks>
+/// <param name="bytes">The encoded keystroke.</param>
+internal sealed class TerminalInputEventArgs(byte[] bytes) : EventArgs
+{
+    public byte[] Bytes { get; } = bytes;
+}

src/DodoSSH.Client.Android/Views/TerminalInputSurface.cs

@@ -0,0 +1,335 @@
+using System.Text;
+
+using Avalonia;
+using Avalonia.Controls;
+using Avalonia.Input;
+using Avalonia.Input.TextInput;
+using Avalonia.Media;
+
+namespace DodoSSH.Client.Android.Views;
+
+/// <summary>
+/// The terminal's keyboard, as an Avalonia control that draws nothing.
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>This exists so that the software keyboard is configured by us rather than by Chromium</b>, and that
+/// single sentence is the whole of ADR 0015. A phone keyboard over a terminal has to be told to stop
+/// autocorrecting, stop predicting, stop capitalising and stop offering a suggestion strip, and on Android
+/// the only setting that actually achieves it is an <c>EditorInfo.inputType</c> carrying
+/// <c>TYPE_TEXT_VARIATION_VISIBLE_PASSWORD</c>. A page inside a <c>WebView</c> cannot reach that field —
+/// Chromium derives the whole <c>EditorInfo</c> from the focused DOM node, and xterm.js already sets every
+/// attribute the web platform offers (<c>autocorrect</c>, <c>autocapitalize</c>, <c>spellcheck</c>) to no
+/// effect, because they are advisory and Gboard ignores them.
+/// </para>
+/// <para>
+/// <b>Avalonia can reach it, on the version this repository already pins.</b>
+/// <c>AndroidInputMethod.OnCreateInputConnection</c> in Avalonia.Android 12.1.1 reads:
+/// <code>
+/// if ((int)outAttrs.InputType == 1 &amp;&amp; options.ShowSuggestions == false)
+///     outAttrs.InputType = (InputTypes)(outAttrs.InputType | 0x80080);
+/// </code>
+/// <c>0x80080</c> is <c>TYPE_TEXT_VARIATION_VISIBLE_PASSWORD | TYPE_TEXT_FLAG_NO_SUGGESTIONS</c> — the exact
+/// pair Termux's own <c>TerminalView</c> sets, and the reason typing in Termux feels like a terminal. The
+/// guard is why the constructor leaves <c>ContentType</c>, <c>Multiline</c> and <c>AutoCapitalization</c>
+/// alone: any of those would move <c>InputType</c> off <c>1</c> and the branch would never be taken.
+/// <c>ShowSuggestions</c> is a <c>bool?</c> defaulting to <c>null</c>, so it must be set to <c>false</c>
+/// rather than merely left.
+/// </para>
+/// <para>
+/// <b>Two options are set, and the second one is not optional either.</b> The default
+/// <c>ReturnKeyType</c> maps to <c>IME_ACTION_DONE</c> on a non-multiline field, which closed the keyboard
+/// on every Enter — that is, on every command. <c>ReturnKeyType.Return</c> is <c>IME_ACTION_NONE</c> and
+/// touches nothing else. The constructor carries the full argument, including why the obvious alternative
+/// of <c>Multiline</c> cannot be used without giving the suggestion strip back.
+/// </para>
+/// <para>
+/// <b>One pixel, and never drawn.</b> Focus in Avalonia has nothing to do with hit-testing, so the input
+/// target does not need to cover the terminal — and must not, because the renderer is a native child view
+/// composited above everything Avalonia paints, so a control stretched over it would be invisible, unable
+/// to receive the touches the <c>WebView</c> consumes first, and in the way of the connect page that shares
+/// this Panel when no session is open. A 1×1 control in the corner is focusable, costs one layout slot, and
+/// is the whole of what an input target has to be. Taps on the terminal itself are noticed by
+/// <see cref="Platform.RendererFocus"/> instead, which sees them where they actually land.
+/// </para>
+/// <para>
+/// <b>Nothing is retained.</b> <see cref="TerminalImeClient"/> reports no surrounding text and no preedit,
+/// because a terminal has no document for an input method to edit — the remote holds the line and the only
+/// honest answer to "what is around the cursor" is "ask the host, which you cannot". Committed text is
+/// turned into bytes and forwarded the instant it arrives.
+/// </para>
+/// </remarks>
+internal sealed class TerminalInputSurface : Control
+{
+    /// <summary>
+    /// Bytes the shell should receive, already encoded.
+    /// </summary>
+    /// <remarks>
+    /// Raised rather than sent here, because the Ctrl latch belongs to the accessory row and the session to
+    /// the view model. <c>TerminalScreen.SendAsync</c> owns both and applies the latch to anything one byte
+    /// long, which is exactly the set a latch should reach: a typed letter, yes; a pasted line or a cursor
+    /// key's escape sequence, no.
+    /// </remarks>
+    internal event EventHandler<TerminalInputEventArgs>? Input;
+
+    private readonly TerminalImeClient client;
+
+    /// <summary>
+    /// Whether the text event about to arrive has already been sent as a control byte.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// <b>Necessary because handling a key event does not stop the text that follows it.</b> Avalonia's
+    /// <c>AndroidKeyboardEventsHelper</c> raises the key event and then, independently, raises a text input
+    /// event for any key whose <c>UnicodeChar</c> is 32 or above — it reads the handled flag from the first
+    /// only to combine into its own return value, never to skip the second. So Ctrl-C on a hardware keyboard
+    /// produces both a <c>KeyDown</c> this class turns into <c>0x03</c> and a text input carrying a literal
+    /// <c>c</c>, and without this flag the shell receives an interrupt followed by a stray character.
+    /// </para>
+    /// <para>
+    /// A plain field rather than anything cleverer because the two events are raised one after the other in
+    /// the same dispatch, and it is cleared at the top of every <see cref="OnKeyDown"/> so a chord that
+    /// somehow produced no text cannot swallow the next thing typed.
+    /// </para>
+    /// <para>
+    /// The keys in <see cref="TryTranslate"/> need no such guard — Enter, Tab, Escape and Backspace all
+    /// carry a <c>UnicodeChar</c> below 32, so no text event is raised for them in the first place.
+    /// </para>
+    /// </remarks>
+    private bool controlByteAlreadySent;
+
+    public TerminalInputSurface()
+    {
+        client = new TerminalImeClient(this);
+
+        Focusable = true;
+
+        // Drawn nowhere and hit-tested never — see the remarks. The size is 1 rather than 0 only because a
+        // zero-area control is a shape no layout pass has a reason to keep.
+        Width = 1;
+        Height = 1;
+        HorizontalAlignment = Avalonia.Layout.HorizontalAlignment.Left;
+        VerticalAlignment = Avalonia.Layout.VerticalAlignment.Top;
+        IsHitTestVisible = false;
+
+        // The one line that fixes the keyboard. See the remarks for why almost nothing else is set.
+        TextInputOptions.SetShowSuggestions(this, false);
+
+        // ◆ AND THE ONE THAT KEEPS THE KEYBOARD UP WHEN A COMMAND IS RUN.
+        //
+        // Avalonia maps ReturnKeyType.Default, on a field that is not Multiline, to IME_ACTION_DONE — and
+        // Done means done: the keyboard performs the action and dismisses itself. Every Enter therefore
+        // closed the keyboard, which on a terminal is every command.
+        //
+        // Return is the member that maps to IME_ACTION_NONE, leaving the key as a plain newline with no
+        // action attached and no reason for the IME to hide. Its whole switch arm is
+        // `returnKeyType - 1 => 0 => 1073741824`, which is 0x40000000, IME_ACTION_NONE.
+        //
+        // Multiline=true reaches the same IME_ACTION_NONE through the default arm and is the wrong way to
+        // get there: it also ORs TYPE_TEXT_FLAG_MULTI_LINE into InputType, which moves InputType off 1 and
+        // silently fails the ShowSuggestions guard above. The two options cannot both be had that way — one
+        // trades the suggestion strip back for a working Enter. ReturnKeyType costs nothing and is the
+        // reason this file sets it rather than Multiline.
+        TextInputOptions.SetReturnKeyType(this, TextInputReturnKeyType.Return);
+
+        AddHandler(TextInputMethodClientRequestedEvent, OnTextInputMethodClientRequested);
+    }
+
+    /// <summary>
+    /// Takes focus and asks for the keyboard.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// The request is made even when focus was already here, which is the case that matters: the system back
+    /// gesture puts the keyboard away without moving focus, so the way back has to be an activation request
+    /// rather than a focus change that would be a no-op. <c>AndroidInputMethod</c> answers
+    /// <c>InputPaneActivationRequested</c> with <c>ShowSoftInput</c>, which is precisely this.
+    /// </para>
+    /// <para>
+    /// Focusing for the first time raises the keyboard on its own — <c>SetClient</c> calls
+    /// <c>RestartInput</c> and <c>ShowSoftInput</c> when it takes a non-null client — so the ordering here is
+    /// belt and braces rather than two mechanisms disagreeing.
+    /// </para>
+    /// </remarks>
+    internal void Acquire()
+    {
+        Focus();
+        client.RequestKeyboard();
+    }
+
+    private void OnTextInputMethodClientRequested(object? sender, TextInputMethodClientRequestedEventArgs e) =>
+        e.Client = client;
+
+    /// <remarks>
+    /// Everything the IME commits and everything a hardware keyboard types that is not in
+    /// <see cref="TryTranslate"/>'s table. UTF-8 because that is what the pty is told the session is, and
+    /// because a multi-byte character has to arrive as one write or a partial sequence reaches the remote.
+    /// </remarks>
+    protected override void OnTextInput(TextInputEventArgs e)
+    {
+        if (controlByteAlreadySent)
+        {
+            controlByteAlreadySent = false;
+            e.Handled = true;
+
+            return;
+        }
+
+        if (string.IsNullOrEmpty(e.Text))
+        {
+            return;
+        }
+
+        // Newlines arrive as carriage returns, whichever way they came. A keyboard with no editor action
+        // attached may commit "\n" as ordinary text rather than sending KEYCODE_ENTER, and a paste carries
+        // whatever the clipboard held — but a terminal's line terminator from a keyboard is CR, and a shell
+        // handed LF for Enter is a shell that looks like it ignored the key. Normalised rather than special
+        // -cased so that a pasted "\r\n" does not run the line twice.
+        var text = e.Text.ReplaceLineEndings("\r");
+
+        Input?.Invoke(this, new TerminalInputEventArgs(Encoding.UTF8.GetBytes(text)));
+        e.Handled = true;
+
+        base.OnTextInput(e);
+    }
+
+    /// <remarks>
+    /// <para>
+    /// Only the keys that have no text of their own are handled here; everything printable is left alone so
+    /// that it arrives once, through <see cref="OnTextInput"/>. Handling both would double every letter.
+    /// </para>
+    /// <para>
+    /// Ctrl with a letter is translated here rather than left to the latch, because a hardware keyboard can
+    /// hold a modifier and this is the only place that sees the chord.
+    /// </para>
+    /// </remarks>
+    protected override void OnKeyDown(KeyEventArgs e)
+    {
+        controlByteAlreadySent = false;
+
+        if (e.KeyModifiers.HasFlag(KeyModifiers.Control) && TryControlChord(e.Key, out var chord))
+        {
+            Input?.Invoke(this, new TerminalInputEventArgs(chord));
+
+            // The letter's own text event is still coming; see the field's remarks.
+            controlByteAlreadySent = true;
+            e.Handled = true;
+
+            return;
+        }
+
+        if (TryTranslate(e.Key, out var bytes))
+        {
+            Input?.Invoke(this, new TerminalInputEventArgs(bytes));
+            e.Handled = true;
+
+            return;
+        }
+
+        base.OnKeyDown(e);
+    }
+
+    /// <remarks>
+    /// The ASCII rule and nothing cleverer, matching <c>TerminalScreen.SendAsync</c>'s latch: Ctrl-A through
+    /// Ctrl-Z are <c>0x01</c>–<c>0x1A</c>, which is the letter masked with <c>0x1F</c>.
+    /// </remarks>
+    private static bool TryControlChord(Key key, out byte[] bytes)
+    {
+        if (key is >= Key.A and <= Key.Z)
+        {
+            bytes = [(byte)((byte)('a' + (key - Key.A)) & 0x1F)];
+
+            return true;
+        }
+
+        bytes = [];
+
+        return false;
+    }
+
+    /// <summary>
+    /// The keys a terminal expects as escape sequences rather than as text.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// Cursor keys are written as CSI rather than SS3, which is the same choice — and the same reasoning —
+    /// as the accessory row's table in <c>TerminalScreen</c>: the difference between <c>CSI A</c> and
+    /// <c>SS3 A</c> is the difference between working in bash and not, and nothing here tracks the remote's
+    /// application-cursor mode.
+    /// </para>
+    /// <para>
+    /// <b><see cref="Key.Delete"/> is deliberately spent on backspace, and it is the one lossy choice in
+    /// this file.</b> Avalonia's Android input connection turns an IME's <c>deleteSurroundingText</c> into a
+    /// dispatched <c>KEYCODE_FORWARD_DEL</c> — see <c>TextEditBuffer.Remove</c> — so a soft keyboard that
+    /// deletes through the edit buffer rather than through a real <c>KEYCODE_DEL</c> arrives here as
+    /// <c>Delete</c>. Mapping it to <c>CSI 3 ~</c> would be correct for a hardware forward-delete and would
+    /// make backspace delete forwards on any keyboard taking the other route, which is the worse failure by
+    /// a wide margin: a phone has no forward-delete key to miss, and a broken backspace makes the terminal
+    /// unusable. Revisit if a hardware keyboard's Delete is ever worth more than this insurance.
+    /// </para>
+    /// </remarks>
+    private static bool TryTranslate(Key key, out byte[] bytes)
+    {
+        bytes = key switch
+        {
+            Key.Enter => [0x0D],
+            Key.Tab => [0x09],
+            Key.Escape => [0x1B],
+            Key.Back or Key.Delete => [0x7F],
+
+            Key.Up => [0x1B, (byte)'[', (byte)'A'],
+            Key.Down => [0x1B, (byte)'[', (byte)'B'],
+            Key.Right => [0x1B, (byte)'[', (byte)'C'],
+            Key.Left => [0x1B, (byte)'[', (byte)'D'],
+
+            Key.Home => [0x1B, (byte)'[', (byte)'H'],
+            Key.End => [0x1B, (byte)'[', (byte)'F'],
+            Key.Insert => [0x1B, (byte)'[', (byte)'2', (byte)'~'],
+            Key.PageUp => [0x1B, (byte)'[', (byte)'5', (byte)'~'],
+            Key.PageDown => [0x1B, (byte)'[', (byte)'6', (byte)'~'],
+
+            _ => [],
+        };
+
+        return bytes.Length > 0;
+    }
+
+    /// <summary>
+    /// What Android's input method is told about the text it is editing, which is that there is none.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// <c>SupportsSurroundingText</c> and <c>SupportsPreedit</c> are both false, and
+    /// <see cref="SurroundingText"/> is empty, because they are the truth: the line being edited lives on
+    /// the far side of an SSH channel and this client has never been told what is on it. Claiming otherwise
+    /// would invite the IME to reason about text it cannot see and to issue edits against positions that
+    /// mean nothing here.
+    /// </para>
+    /// <para>
+    /// <see cref="Selection"/> has to be settable because <c>TextEditBuffer</c> writes it before every
+    /// replacement, so it is stored and ignored rather than refused — a property that threw would take the
+    /// input connection down with it.
+    /// </para>
+    /// <para>
+    /// <see cref="CursorRectangle"/> is empty for the same reason the control is 1×1: it positions an IME's
+    /// composition popup, and a field that never composes has no popup to place.
+    /// </para>
+    /// </remarks>
+    private sealed class TerminalImeClient(Visual visual) : TextInputMethodClient
+    {
+        public override Visual TextViewVisual => visual;
+
+        public override bool SupportsPreedit => false;
+
+        public override bool SupportsSurroundingText => false;
+
+        public override string SurroundingText => string.Empty;
+
+        public override Rect CursorRectangle => default;
+
+        public override TextSelection Selection { get; set; }
+
+        /// <remarks>Exposed because <c>RaiseInputPaneActivationRequested</c> is protected on the base.</remarks>
+        internal void RequestKeyboard() => RaiseInputPaneActivationRequested();
+    }
+}

src/DodoSSH.Client.Android/Views/TerminalScreen.axaml

@@ -19,9 +19,13 @@
 
     ── the screen a shell gets ───────────────────────────────────────────────────────────────────────────
     A connected phone shows one bar and then the terminal. The vault header, the shells strip and the
-    four-entry bottom bar are all collapsed by PhoneShell while this surface is up, and what replaces them
+    three-entry bottom bar are all collapsed by PhoneShell while a shell is showing, and what replaces them
     is the row below: back, the sessions, and the way to open another one.
 
+    With nothing running it is only the header and the strip. This surface is then the connect page below —
+    a box and the machines connected to before — which is not what the screen was being cleared for, and is
+    the one screen reachable by closing your last tab, so the nav stays. See PhoneShell.RefreshChrome.
+
     That is a trade, and the thing bought is the only one a terminal really wants. At 360dp the chrome this
     screen used to sit inside came to 254 pixels of a roughly 780-pixel display — a third of it — and every
     one of those rows was about somewhere the user was not. What is given up is the bottom bar's one-tap
@@ -29,9 +33,94 @@
     gesture does the same thing the arrow does.
   -->
 
+  <UserControl.Styles>
+
+    <!--
+      ── the connecting step list ─────────────────────────────────────────────────────────────────────
+      The same five rows the desktop's ConnectingCard draws, from the same reported phases, in this head's
+      own sizes. Kept here rather than in Phone.axaml because nothing else on this head has a step list —
+      the theme file is for what more than one screen shares, and a rule that exists for one control is
+      easier to read beside it.
+
+      Amber for the step in flight, green behind it, red where it stopped. That is the palette's rule
+      rather than an exception to it: green is what is true and purple is what you can press, and a step
+      still happening is neither. See ConnectingCard.axaml for the longer version of this argument, and
+      Palette.axaml for the rule itself.
+
+      A phone needs this more than a desktop does, which is the same thing the connecting block below
+      already says about itself: mobile links are slower and drop more often, so the stretch this describes
+      is longer here and more likely to end badly.
+    -->
+    <!--
+      Its own FontFamily rather than the row also carrying the mono class, which is this head's convention
+      and not a stylistic preference: Phone.axaml's mono sets a colour and a size along with the family, so
+      a caption wearing both classes would be asking two rules for one Foreground and settling it on style
+      ordering. Every other text class here — body, label, title, detail — names its own family for exactly
+      that reason. The desktop's mono sets the family alone, which is why ConnectingCard composes the two
+      and this does not.
+    -->
+    <Style Selector="TextBlock.stepcaption">
+      <Setter Property="FontFamily" Value="{StaticResource MonoFont}" />
+      <Setter Property="Foreground" Value="{StaticResource TextFaint}" />
+      <Setter Property="FontSize" Value="11" />
+      <Setter Property="VerticalAlignment" Value="Center" />
+    </Style>
+    <Style Selector="TextBlock.stepcaption.done">
+      <Setter Property="Foreground" Value="{StaticResource TextDim}" />
+    </Style>
+    <Style Selector="TextBlock.stepcaption.running">
+      <Setter Property="Foreground" Value="{StaticResource WarnText}" />
+      <Setter Property="FontWeight" Value="SemiBold" />
+    </Style>
+    <Style Selector="TextBlock.stepcaption.stopped">
+      <Setter Property="Foreground" Value="{StaticResource DangerText}" />
+    </Style>
+
+    <!-- Fixed width and centred: four different characters on a ragged edge is a list that looks broken. -->
+    <Style Selector="TextBlock.stepmark">
+      <Setter Property="Foreground" Value="{StaticResource BorderMid}" />
+      <Setter Property="FontSize" Value="11" />
+      <Setter Property="Width" Value="13" />
+      <Setter Property="TextAlignment" Value="Center" />
+      <Setter Property="VerticalAlignment" Value="Center" />
+    </Style>
+    <Style Selector="TextBlock.stepmark.done">
+      <Setter Property="Foreground" Value="{StaticResource Live}" />
+    </Style>
+    <Style Selector="TextBlock.stepmark.running">
+      <Setter Property="Foreground" Value="{StaticResource Warn}" />
+    </Style>
+    <Style Selector="TextBlock.stepmark.stopped">
+      <Setter Property="Foreground" Value="{StaticResource Danger}" />
+    </Style>
+
+    <!--
+      4 rather than the desktop's 5, which is the only deliberate difference between the two heads here:
+      this bar sits in a column 24 from each edge of a 360dp screen rather than under a 460-wide card, so
+      the same height reads as a heavier rule across a narrower span.
+    -->
+    <Style Selector="ProgressBar.steptrack">
+      <Setter Property="Height" Value="4" />
+      <Setter Property="MinHeight" Value="4" />
+      <Setter Property="CornerRadius" Value="2" />
+      <Setter Property="Background" Value="{StaticResource Chip}" />
+      <Setter Property="Foreground" Value="{StaticResource Warn}" />
+    </Style>
+    <Style Selector="ProgressBar.steptrack.stopped">
+      <Setter Property="Foreground" Value="{StaticResource Danger}" />
+    </Style>
+
+  </UserControl.Styles>
+
   <Panel>
 
-  <Grid RowDefinitions="Auto,*,Auto">
+  <!--
+    Four rows since the disconnect strip arrived. It is Auto and empty except while the selected tab's shell
+    has stopped, so an ordinary session pays nothing for it — and when it is up it takes its height off the
+    renderer rather than drawing over it, which on this head is not optional either: the renderer is a native
+    child view composited above everything Avalonia draws, so a banner over its rectangle would be invisible.
+  -->
+  <Grid RowDefinitions="Auto,*,Auto,Auto">
 
     <!-- ============ the bar ============ -->
     <!--
@@ -45,7 +134,10 @@
       to 34, the pills to 30 — because a bar that shrank around controls that did not would only have moved
       the clipping somewhere harder to see.
     -->
-    <Border Grid.Row="0" Height="35" Background="{StaticResource Chrome}"
+    <!-- DeepChrome rather than Chrome, since v5: this bar is what the vault header, the strip and the
+         bottom bar collapse into while a shell is showing, so it takes their surface along with their
+         job — see the remark on the header in PhoneShell.axaml. -->
+    <Border Grid.Row="0" Height="35" Background="{StaticResource DeepChrome}"
             BorderBrush="{StaticResource Border}" BorderThickness="0,0,0,1">
       <Grid ColumnDefinitions="Auto,*,Auto">
 
@@ -83,15 +175,21 @@
                   row where nothing sits above or below it to be hit by mistake.
                 -->
                 <Border Background="{StaticResource Panel}" BorderBrush="{StaticResource BorderMid}"
-                        BorderThickness="1" CornerRadius="9" Height="30">
+                        BorderThickness="1" CornerRadius="10" Height="30">
                   <StackPanel Orientation="Horizontal">
-                    <Button Classes="row" MinHeight="28" Padding="11,0" CornerRadius="9"
+                    <Button Classes="row" MinHeight="28" Padding="11,0" CornerRadius="10"
                             VerticalContentAlignment="Center"
                             Command="{Binding $parent[views:TerminalScreen].((vm:MainWindowViewModel)DataContext).SelectTabCommand}"
                             CommandParameter="{Binding}">
                       <StackPanel Orientation="Horizontal" Spacing="7" VerticalAlignment="Center">
-                        <!-- Green only while there is a shell behind it; see the same dot in PhoneShell. -->
-                        <Ellipse Classes="dot" Classes.live="{Binding IsLive}" Width="6" Height="6"
+                        <!--
+                          Green only while there is a shell behind it, amber while one is being made, and
+                          red where a connection broke rather than a shell exiting; see the same dot in
+                          PhoneShell, and Phone.axaml for why neither of the last two is a rule broken.
+                        -->
+                        <Ellipse Classes="dot" Classes.live="{Binding IsLive}"
+                                 Classes.connecting="{Binding IsConnecting}"
+                                 Classes.dropped="{Binding IsDropped}" Width="6" Height="6"
                                  VerticalAlignment="Center" />
                         <TextBlock Classes="mono" FontSize="12" FontWeight="SemiBold"
                                    Text="{Binding Label}" />
@@ -112,7 +210,7 @@
                       It reads as a misprint, which for the control that ends a session is the wrong thing
                       to look like.
                     -->
-                    <Button Classes="row" MinHeight="28" Width="44" Padding="0" CornerRadius="0,9,9,0"
+                    <Button Classes="row" MinHeight="28" Width="44" Padding="0" CornerRadius="0,10,10,0"
                             HorizontalContentAlignment="Center" VerticalContentAlignment="Center"
                             BorderBrush="{StaticResource BorderMid}" BorderThickness="1,0,0,0"
                             Command="{Binding $parent[views:TerminalScreen].((vm:MainWindowViewModel)DataContext).CloseTabCommand}"
@@ -243,7 +341,7 @@
                         <TextBlock Classes="mono" FontSize="13" FontWeight="SemiBold"
                                    Text="{Binding HostLabel}" TextTrimming="CharacterEllipsis" />
                         <TextBlock Classes="detail" FontSize="10.5" Text="{Binding Address}"
-                                   TextTrimming="CharacterEllipsis" />
+                                   TextWrapping="NoWrap" TextTrimming="CharacterEllipsis" />
                       </StackPanel>
                       <TextBlock Grid.Column="1" Classes="detail" FontSize="10"
                                  VerticalAlignment="Center" Text="{Binding Started}" />
@@ -277,11 +375,82 @@
         <TextBlock Classes="title" FontSize="13" Text="{Binding SelectedTab.Label}" />
         <TextBlock Classes="detail" FontSize="11" Foreground="{StaticResource TextDim}"
                    TextWrapping="Wrap" Text="{Binding SelectedTab.Address}" />
-        <TextBlock Classes="body" Text="{Binding SelectedTab.Status}" />
-        <Button Classes="row" MinHeight="44" Padding="14,0" HorizontalAlignment="Left"
-                Command="{Binding CloseTabCommand}" CommandParameter="{Binding SelectedTab}">
-          <TextBlock Classes="label" FontSize="9" Text="CLOSE THIS TAB" />
-        </Button>
+
+        <!--
+          Where a single unchanging "connecting…" used to be. The track counts steps that really finished
+          against the five there are — StepsDone over StepCount, never a percentage, because the arithmetic
+          that makes a percentage is the arithmetic that starts inventing one. See TerminalTabViewModel.
+
+          Drawn for both states rather than once per state: a refused connection has the same five rows and
+          the same track, and the only differences are that one row is red and the track stops where it got
+          to. Two templates kept identical for the sake of a colour is how the two drift apart.
+        -->
+        <ProgressBar Classes="steptrack" Classes.stopped="{Binding SelectedTab.IsFailed}"
+                     Minimum="0" Maximum="{Binding SelectedTab.StepCount}"
+                     Value="{Binding SelectedTab.StepsDone, Mode=OneWay}" />
+
+        <ItemsControl ItemsSource="{Binding SelectedTab.Steps}">
+          <ItemsControl.ItemsPanel>
+            <ItemsPanelTemplate>
+              <StackPanel Spacing="6" />
+            </ItemsPanelTemplate>
+          </ItemsControl.ItemsPanel>
+          <ItemsControl.ItemTemplate>
+            <DataTemplate x:DataType="vm:ConnectionStepViewModel">
+              <StackPanel Orientation="Horizontal" Spacing="9">
+                <TextBlock Classes="stepmark"
+                           Classes.done="{Binding IsDone}"
+                           Classes.running="{Binding IsRunning}"
+                           Classes.stopped="{Binding IsStopped}"
+                           Text="{Binding Mark}" />
+                <TextBlock Classes="stepcaption"
+                           Classes.done="{Binding IsDone}"
+                           Classes.running="{Binding IsRunning}"
+                           Classes.stopped="{Binding IsStopped}"
+                           Text="{Binding Caption}" />
+              </StackPanel>
+            </DataTemplate>
+          </ItemsControl.ItemTemplate>
+        </ItemsControl>
+
+        <!--
+          Only for a refusal now. While a connection is being made this used to be the whole of what this
+          screen said, and it is now the step list's running row said twice — so it is shown for the one
+          state the list cannot put into words: why it stopped.
+        -->
+        <TextBlock Classes="body" Text="{Binding SelectedTab.Status}"
+                   Foreground="{StaticResource Danger}"
+                   IsVisible="{Binding SelectedTab.IsFailed}" />
+
+        <!--
+          Two 44-high targets side by side rather than one, and the second is the logs: the step list is
+          this attempt and the log is every other one, which is the question a connection that is taking too
+          long on a mobile link actually raises — has this machine ever worked from here. Reached the
+          ordinary way, through ShowScreenCommand, exactly as the rail and MORE reach it.
+        -->
+        <StackPanel Orientation="Horizontal" Spacing="8" HorizontalAlignment="Left">
+          <!--
+            Only for a refusal, and it runs the same command the disconnect strip below the renderer does —
+            the vault's own connect path, back into this same tab. Worth more on a phone than on a desk: the
+            commonest refusal here is a mobile link that was not up yet, and the alternative is going back
+            to the hosts screen to find a machine whose name is already on this card.
+          -->
+          <Button Classes="row" MinHeight="44" Padding="14,0"
+                  Command="{Binding ReconnectTabCommand}" CommandParameter="{Binding SelectedTab}"
+                  IsEnabled="{Binding CanReconnectSelectedTab, FallbackValue=False}"
+                  IsVisible="{Binding SelectedTab.IsFailed}">
+            <TextBlock Classes="label" FontSize="9" Text="TRY AGAIN" />
+          </Button>
+          <Button Classes="row" MinHeight="44" Padding="14,0"
+                  Command="{Binding CloseTabCommand}" CommandParameter="{Binding SelectedTab}">
+            <TextBlock Classes="label" FontSize="9" Text="CLOSE THIS TAB" />
+          </Button>
+          <Button Classes="row" MinHeight="44" Padding="14,0"
+                  Command="{Binding ShowScreenCommand}"
+                  CommandParameter="{x:Static vm:ShellScreen.Logs}">
+            <TextBlock Classes="label" FontSize="9" Text="SHOW LOGS" />
+          </Button>
+        </StackPanel>
       </StackPanel>
 
       <!--
@@ -302,6 +471,29 @@
         design's #171A26. See WebAssets/terminal.css.
       -->
       <NativeWebView x:Name="Renderer" IsVisible="{Binding IsTerminalShowing, FallbackValue=False}" />
+
+      <!--
+        ============ the keyboard ============
+
+        One pixel in the corner, drawn never, and it is where every character typed into a shell on this
+        phone arrives. ADR 0015 has the whole argument; the short of it is that a software keyboard over a
+        terminal has to be told to stop autocorrecting, and on Android the only setting that works is an
+        inputType on a native input connection — which a page inside a WebView cannot reach and an Avalonia
+        control can, on the version already pinned here.
+
+        Not stretched over the renderer, which would be the obvious shape and is wrong three times: the
+        WebView composites above everything Avalonia paints so it would be invisible, it would never see the
+        touches the WebView consumes first, and it would sit over the connect page that shares this Panel
+        whenever no session is open. Focus does not need area. Taps on the terminal are noticed by
+        RendererFocus, where they actually land.
+
+        Collapsed with the renderer, on the same binding and for a reason of its own: a focusable control is
+        a tab stop and a focus candidate wherever it lives, and this Panel is also the connect page when no
+        session is open. Left always-visible it would be a stray target on a screen with a real text box on
+        it, offering a keyboard configured for a shell to somebody typing a hostname.
+      -->
+      <views:TerminalInputSurface x:Name="Input"
+                                  IsVisible="{Binding IsTerminalShowing, FallbackValue=False}" />
     </Panel>
 
     <!-- ============ accessory keys ============ -->
@@ -313,70 +505,164 @@
     -->
     <!-- Only with a pane to type into: the keys send bytes at a session, and a connecting tab has none. -->
     <!--
-      33 tall rather than 50, a third off it like the bar at the top, with the keys inside coming down from
-      38 to 30 for the reason the pills did: a row that shrank around its own contents would clip them.
+      ◆ TWO FIXED ROWS THAT FILL THE WIDTH, AND NOTHING HERE SCROLLS. THAT IS THE WHOLE CHANGE.
 
-      The five pixels above it are a gap and not a border. The renderer is a native child view, so nothing
-      Avalonia draws can sit on top of it — a hairline between the two would have to be a row of its own —
-      and a terminal whose last line of output is flush against a row of grey keys reads as one surface
-      that has gone wrong rather than as two that are different things.
+      What this replaced was one 33-pixel row holding eleven 30-by-44 keys inside a horizontal ScrollViewer,
+      and it was wrong in two ways that compounded each other.
+
+      ── the scroller was eating the presses ──────────────────────────────────────────────────────────────
+      A horizontally scrollable strip and a row of buttons want opposite things from the same gesture. Every
+      tap starts as a press and a few pixels of travel, and a ScrollViewer that can pan claims exactly that:
+      a thumb coming down at an angle — which is how a thumb comes down — is read as the start of a pan, the
+      button never sees a click, and the row slides slightly instead. Pressing a key and having the keys
+      move is the single most confidence-destroying thing a keyboard can do, and it happened often enough
+      to be the first thing anybody noticed about this row.
+
+      ── and the keys were too small to hit anyway ─────────────────────────────────────────────────────────
+      30 tall. Android's own minimum touch target is 48 and every accessory row in every terminal on the
+      platform is somewhere near it. The height had been taken down from 38 to save a strip of terminal,
+      which is a trade worth making right up until the control stops working, and 30 is past that.
+
+      ── what the other clients do, because all of them agree ─────────────────────────────────────────────
+      Termux, Termius, JuiceSSH, ConnectBot and Blink all ship this row and not one of them scrolls it.
+      They divide the width evenly between a fixed set of keys and, where the set does not fit, they use a
+      second row — Termux ships two by default. That is not convention for its own sake: a key whose
+      position is fixed can be hit without looking, which is the entire point of a modifier row you use
+      while reading the output above it. A key that might have scrolled has to be found first.
+
+      So: two rows, every key stretched to an equal share of the width, none of it scrollable, and 40 tall
+      instead of 30. Nothing was dropped — all eleven keys are here, and the four that mattered most are the
+      ones that benefit, because the arrows were the keys living past the right edge.
+
+      ── what it costs, stated rather than buried ─────────────────────────────────────────────────────────
+      About 51 pixels of terminal, which is four or five lines of output. That is a real price and it is the
+      right one: the row that was cheaper did not work, and a terminal with five more lines and no reachable
+      Ctrl is not the better surface.
+
+      ── the split ────────────────────────────────────────────────────────────────────────────────────────
+      Modifiers and punctuation on top, cursor keys underneath. The arrows go on the lower row because they
+      are the keys pressed repeatedly and in sequence — history, then along the line — so they want the row
+      the thumb rests nearest, and there are only four of them, which leaves them the widest targets on the
+      whole keypad. The text-size buttons ride at that row's right-hand end, outside the flex.
+
+      The five pixels above the whole thing are a gap and not a border. The renderer is a native child view,
+      so nothing Avalonia draws can sit on top of it — a hairline between the two would have to be a row of
+      its own — and a terminal whose last line of output is flush against a row of grey keys reads as one
+      surface that has gone wrong rather than as two that are different things.
     -->
-    <Border Grid.Row="2" IsVisible="{Binding IsTerminalShowing}" Height="33" Margin="0,5,0,0">
-      <Grid ColumnDefinitions="*,Auto">
+    <!-- ============ what a terminal says once its shell has stopped ============ -->
+    <!--
+      The phone's half of SessionEndedBar, and it needs it more than the desk does: a mobile link drops on
+      its own, the screen goes off, and the tab that went quiet is discovered later with no idea whether the
+      shell was exited or cut off. So it names which, prints the sentence that came with it, and offers the
+      one useful answer.
 
-        <ScrollViewer Grid.Column="0" HorizontalScrollBarVisibility="Auto"
-                      VerticalScrollBarVisibility="Disabled">
-          <StackPanel x:Name="AccessoryKeys" Orientation="Horizontal" Spacing="5" Margin="12,0"
-                      VerticalAlignment="Center" />
-        </ScrollViewer>
+      Two lines rather than the desktop's one, because 360dp cannot hold a word, a sentence and two buttons
+      across. The word and the buttons share the top line — the buttons are where a thumb already is — and
+      the sentence wraps underneath at two lines maximum, which is enough for anything sshd or a socket
+      error actually says and short of a strip that eats the terminal.
+
+      RECONNECT is disabled rather than hidden when there is nothing to dial: a machine typed in by hand has
+      no keychain entry to redial, and a control visibly refusing answers "why not" better than an absent one.
+      See MainWindowViewModel.CanReconnectSelectedTab.
+    -->
+    <Border Grid.Row="2" Background="{StaticResource DeepChrome}"
+            BorderBrush="{StaticResource Border}" BorderThickness="0,1,0,0" Padding="16,8"
+            IsVisible="{Binding SessionEndedText, Converter={x:Static StringConverters.IsNotNullOrEmpty}, FallbackValue=False}">
+      <StackPanel Spacing="6">
+
+        <Grid ColumnDefinitions="*,Auto">
+          <StackPanel Grid.Column="0" Orientation="Horizontal" Spacing="7" VerticalAlignment="Center">
+            <Ellipse Classes="dot" Classes.dropped="{Binding IsSessionDropped}" Width="7" Height="7"
+                     VerticalAlignment="Center" />
+            <TextBlock Classes="label" FontSize="9" Text="{Binding SessionEndedText}" />
+          </StackPanel>
+
+          <StackPanel Grid.Column="1" Orientation="Horizontal" Spacing="8">
+            <Button Classes="row" MinHeight="40" Padding="12,0"
+                    Command="{Binding ReconnectTabCommand}" CommandParameter="{Binding SelectedTab}"
+                    IsEnabled="{Binding CanReconnectSelectedTab, FallbackValue=False}">
+              <TextBlock Classes="label" FontSize="9" Text="RECONNECT" />
+            </Button>
+            <Button Classes="row" MinHeight="40" Padding="12,0"
+                    Command="{Binding CloseTabCommand}" CommandParameter="{Binding SelectedTab}">
+              <TextBlock Classes="label" FontSize="9" Text="CLOSE" />
+            </Button>
+          </StackPanel>
+        </Grid>
+
+        <TextBlock Classes="detail" FontSize="10.5" Foreground="{StaticResource TextDim}"
+                   TextWrapping="Wrap" MaxLines="2" TextTrimming="CharacterEllipsis"
+                   Text="{Binding SessionEndedDetail}"
+                   IsVisible="{Binding SessionEndedDetail, Converter={x:Static StringConverters.IsNotNullOrEmpty}}" />
+
+      </StackPanel>
+    </Border>
+
+    <Border Grid.Row="3" IsVisible="{Binding IsTerminalShowing}" Margin="0,5,0,0">
+      <StackPanel Spacing="4">
 
         <!--
-          The text-size buttons, pinned at this row's right-hand end rather than scrolling with the keys
-          beside them.
-
-          They used to have a row of their own above the terminal, on the argument that the accessory row
-          scrolls and these two must never be off-screen — a terminal too small to read is exactly the state
-          in which hunting for the control that fixes it is worst. That argument is answered rather than
-          abandoned: outside the ScrollViewer they cannot scroll away, and the row they had costs 36 pixels
-          on a surface this change exists to give back.
-
-          A phone cannot press Ctrl+plus. The desktop head has that chord and needs no buttons; this head
-          has no keyboard to press it with, which is why the two heads differ here and nowhere else in this
-          screen.
-
-          Disabled at the ends rather than clamping silently. A button that keeps accepting taps and does
-          nothing reads as the terminal having stopped responding, which is the one thing this screen must
-          never look like.
-
-          ◆ Focusable="False", for the reason the accessory keys carry in code — see
-          TerminalScreen.axaml.cs. Every button on this row sits over a terminal somebody is typing into,
-          and an ordinary one takes Avalonia's focus off the NativeWebView on tap, which makes the package
-          resign the page's focus. Making the text smaller must not cost the keyboard.
+          Modifiers and punctuation. Populated from the table in the code-behind, which is also where the
+          bytes each one sends live; a UniformGrid rather than a StackPanel so the seven divide the width
+          between them instead of taking what their labels happen to need — "ctrl" is four characters and
+          "|" is one, and on a row of keys those two want the same target, not proportional ones.
         -->
-        <StackPanel Grid.Column="1" Orientation="Horizontal" Spacing="5" Margin="8,0,12,0"
-                    VerticalAlignment="Center">
-          <Border Width="1" Height="18" Background="{StaticResource Border}" Margin="0,0,3,0"
-                  VerticalAlignment="Center" />
+        <UniformGrid x:Name="AccessoryKeys" Columns="7" Margin="8,0" />
 
-          <Button Classes="row" MinHeight="30" Height="30" MinWidth="40" Padding="0" CornerRadius="9"
-                  HorizontalContentAlignment="Center" VerticalContentAlignment="Center"
-                  Background="{StaticResource Panel}" BorderBrush="{StaticResource BorderMid}"
-                  BorderThickness="1" Focusable="False"
-                  Command="{Binding ShrinkTerminalFontCommand}"
-                  IsEnabled="{Binding CanShrinkTerminalFont}">
-            <TextBlock Classes="mono" FontSize="13" Text="A−" />
-          </Button>
+        <!--
+          The cursor keys, and the two text-size buttons riding at the end.
 
-          <Button Classes="row" MinHeight="30" Height="30" MinWidth="40" Padding="0" CornerRadius="9"
-                  HorizontalContentAlignment="Center" VerticalContentAlignment="Center"
-                  Background="{StaticResource Panel}" BorderBrush="{StaticResource BorderMid}"
-                  BorderThickness="1" Focusable="False"
-                  Command="{Binding EnlargeTerminalFontCommand}"
-                  IsEnabled="{Binding CanEnlargeTerminalFont}">
-            <TextBlock Classes="mono" FontSize="15" Text="A+" />
-          </Button>
-        </StackPanel>
-      </Grid>
+          Grid rather than a second UniformGrid across the whole width: the arrows share what is left after
+          A− and A+ have taken their fixed 44 each, so the four of them flex and the two of them do not.
+          Putting all six in one uniform row would make the font buttons as wide as an arrow key and move
+          them every time the row's contents changed.
+        -->
+        <Grid ColumnDefinitions="*,Auto" Margin="8,0">
+
+          <UniformGrid x:Name="ArrowKeys" Grid.Column="0" Columns="4" />
+
+          <!--
+            The text-size buttons.
+
+            They used to have a row of their own above the terminal, on the argument that the accessory row
+            scrolls and these two must never be off-screen — a terminal too small to read is exactly the
+            state in which hunting for the control that fixes it is worst. Nothing on this keypad scrolls
+            any more, so that argument is now satisfied by construction rather than by placement.
+
+            A phone cannot press Ctrl+plus. The desktop head has that chord and needs no buttons; this head
+            has no keyboard to press it with, which is why the two heads differ here and nowhere else in
+            this screen.
+
+            Disabled at the ends rather than clamping silently. A button that keeps accepting taps and does
+            nothing reads as the terminal having stopped responding, which is the one thing this screen must
+            never look like.
+
+            ◆ Focusable="False", for the reason the accessory keys carry in code — see
+            TerminalScreen.axaml.cs. Every button on this keypad sits over a terminal somebody is typing
+            into, and an ordinary one takes Avalonia's focus off the input surface on tap. Making the text
+            smaller must not cost the keyboard.
+          -->
+          <!-- No Spacing: Button.hotkey carries a 2-pixel side margin, so the gaps here match the
+               UniformGrid cells' without being counted twice. -->
+          <StackPanel Grid.Column="1" Orientation="Horizontal" Margin="8,0,0,0">
+            <Border Width="1" Height="20" Background="{StaticResource Border}" Margin="0,0,4,0"
+                    VerticalAlignment="Center" />
+
+            <Button Classes="hotkey" Width="44" Focusable="False"
+                    Command="{Binding ShrinkTerminalFontCommand}"
+                    IsEnabled="{Binding CanShrinkTerminalFont}">
+              <TextBlock Classes="mono" FontSize="13" Text="A−" />
+            </Button>
+
+            <Button Classes="hotkey" Width="44" Focusable="False"
+                    Command="{Binding EnlargeTerminalFontCommand}"
+                    IsEnabled="{Binding CanEnlargeTerminalFont}">
+              <TextBlock Classes="mono" FontSize="15" Text="A+" />
+            </Button>
+          </StackPanel>
+        </Grid>
+      </StackPanel>
     </Border>
 
   </Grid>
@@ -389,9 +675,10 @@
       of them.
 
       It lives here rather than in PhoneShell for the reason the add sheet lives in its own screen: nothing
-      but this surface raises it. The scrim reaching only the screen area is not a compromise here the way
-      it was there — the bottom bar is collapsed while a terminal is showing, so the screen area is the
-      display.
+      but this surface raises it. The scrim reaching only the screen area is the full display while a shell
+      is showing, because the bottom bar is collapsed for it. Raised from the connect page it is not, and
+      that lands on the add sheet's own answer: the three destinations under the scrim stay tappable, and for
+      a menu of places to connect from, navigating away is a perfectly good way to decide not to.
 
       Every row navigates away from the terminal. That is not a side effect of the menu, it is the menu:
       each of the three destinations is a picker, and the shell they open lands back on this surface as a

src/DodoSSH.Client.Android/Views/TerminalScreen.axaml.cs

@@ -1,12 +1,15 @@
+using System.ComponentModel;
 using System.Text;
 
 using Avalonia;
 using Avalonia.Controls;
+using Avalonia.Controls.Primitives;
 using Avalonia.Interactivity;
 using Avalonia.Layout;
 using Avalonia.Markup.Xaml;
 using Avalonia.Media;
 using Avalonia.Platform;
+using Avalonia.Threading;
 
 using DodoSSH.Client.Android.Platform;
 using DodoSSH.Client.Shell.ViewModels;
@@ -37,7 +40,14 @@ internal sealed partial class TerminalScreen : UserControl
     /// keys to fit 360dp produces ten targets too small to hit.
     /// </para>
     /// </remarks>
-    private static readonly (string Label, byte[] Bytes, bool Latches)[] Keys =
+    /// <summary>The upper row: the modifiers, and the punctuation a symbols layer is two taps away from.</summary>
+    /// <remarks>
+    /// Seven, which is what the row divides between — see the keypad's remark in TerminalScreen.axaml for
+    /// why the set is fixed and flexed rather than scrollable. Adding an eighth is a matter of adding it
+    /// here and changing the <c>Columns</c> on the UniformGrid; the two numbers have to agree, and the row
+    /// simply gets narrower keys.
+    /// </remarks>
+    private static readonly (string Label, byte[] Bytes, bool Latches)[] ModifierKeys =
     [
         ("esc", [0x1B], false),
         ("tab", [0x09], false),
@@ -46,12 +56,31 @@ internal sealed partial class TerminalScreen : UserControl
         ("-", [(byte)'-'], false),
         ("~", [(byte)'~'], false),
         ("/", [(byte)'/'], false),
+    ];
 
-        // The cursor keys, as the ANSI sequences a PTY expects. Written out rather than composed, because
-        // the difference between CSI A and SS3 A is the difference between working in bash and not.
-        ("↑", [0x1B, (byte)'[', (byte)'A'], false),
-        ("↓", [0x1B, (byte)'[', (byte)'B'], false),
+    /// <summary>The lower row: the cursor keys, which is where they became reachable.</summary>
+    /// <remarks>
+    /// <para>
+    /// The ANSI sequences a PTY expects, written out rather than composed, because the difference between
+    /// CSI A and SS3 A is the difference between working in bash and not.
+    /// </para>
+    /// <para>
+    /// ◆ <b>Left, down, up, right</b> — the order every other client on the platform puts them in, and not
+    /// the order they were in here. On one flat row the pairs have to read as pairs: horizontal at the ends
+    /// and vertical in the middle means the two keys that undo each other are always neighbours, and a
+    /// thumb learns the row's shape rather than its labels. Up-down-left-right is the order they are named
+    /// in, which is a different thing from the order they are used in.
+    /// </para>
+    /// <para>
+    /// These four were the keys that used to live past the right-hand edge of the scroller — the most
+    /// pressed keys on the keypad, and the ones you had to go looking for.
+    /// </para>
+    /// </remarks>
+    private static readonly (string Label, byte[] Bytes, bool Latches)[] CursorKeys =
+    [
         ("←", [0x1B, (byte)'[', (byte)'D'], false),
+        ("↓", [0x1B, (byte)'[', (byte)'B'], false),
+        ("↑", [0x1B, (byte)'[', (byte)'A'], false),
         ("→", [0x1B, (byte)'[', (byte)'C'], false),
     ];
 
@@ -67,26 +96,119 @@ internal sealed partial class TerminalScreen : UserControl
 
     private Button? controlKey;
 
+    private readonly TerminalInputSurface input;
+
+    /// <summary>The shell whose <c>PropertyChanged</c> this screen is subscribed to, so it can unsubscribe.</summary>
+    private MainWindowViewModel? subscribed;
+
+    /// <summary>
+    /// What <see cref="MainWindowViewModel.IsTerminalShowing"/> was last seen to be.
+    /// </summary>
+    /// <remarks>
+    /// The keyboard is raised on the transition into the terminal and not on every notification of it.
+    /// <c>IsTerminalShowing</c> is a computed property re-raised by several other setters, so acting on the
+    /// notification rather than the edge would call <c>ShowSoftInput</c> repeatedly for one arrival.
+    /// </remarks>
+    private bool terminalWasShowing;
+
     public TerminalScreen()
     {
         AvaloniaXamlLoader.Load(this);
 
         BuildAccessoryRow();
 
+        input = this.FindControl<TerminalInputSurface>("Input")!;
+        input.Input += (_, e) => SendAsync(e.Bytes);
+
         var renderer = this.FindControl<NativeWebView>("Renderer")!;
 
-        renderer.EnvironmentRequested += OnRendererEnvironmentRequested;
+        renderer.EnvironmentRequested += (sender, e) =>
+        {
+            OnRendererEnvironmentRequested(sender, e);
+
+            // The precise moment the platform WebView is being built, which is the earliest one at which
+            // there is a native view to take the keyboard away from — and the one that fires again if
+            // Android reclaims the renderer and the page comes back. Posted rather than called, because the
+            // child is not in the hierarchy yet while its environment is still being asked for.
+            Dispatcher.UIThread.Post(AdoptRenderer, DispatcherPriority.Loaded);
+        };
 
         // The URL is only known once the loopback listener has bound a port, so it cannot be set in XAML.
         DataContextChanged += (_, _) =>
         {
+            if (subscribed is { } previous)
+            {
+                previous.PropertyChanged -= OnShellPropertyChanged;
+                previous.TerminalFocusRequested -= OnTerminalFocusRequested;
+                subscribed = null;
+            }
+
             if (DataContext is MainWindowViewModel shell)
             {
                 renderer.Source = shell.TerminalPageUrl;
+
+                shell.PropertyChanged += OnShellPropertyChanged;
+
+                // The snippets screen's "TYPE INTO {tab}" lands here as well as on the desktop. It types at
+                // a prompt the user is about to want to keep typing at, so the keyboard should be up when
+                // they look. See MainWindowViewModel.TerminalFocusRequested.
+                shell.TerminalFocusRequested += OnTerminalFocusRequested;
+                subscribed = shell;
+
+                terminalWasShowing = shell.IsTerminalShowing;
             }
         };
     }
 
+    /// <summary>
+    /// Raises the keyboard when a shell arrives, and never while one is merely being redrawn.
+    /// </summary>
+    private void OnShellPropertyChanged(object? sender, PropertyChangedEventArgs e)
+    {
+        if (e.PropertyName is not nameof(MainWindowViewModel.IsTerminalShowing) ||
+            sender is not MainWindowViewModel shell)
+        {
+            return;
+        }
+
+        var showing = shell.IsTerminalShowing;
+
+        if (showing == terminalWasShowing)
+        {
+            return;
+        }
+
+        terminalWasShowing = showing;
+
+        if (!showing)
+        {
+            return;
+        }
+
+        // Posted to Loaded so the renderer's own native child, which becomes visible on this same change,
+        // exists by the time AdoptRenderer goes looking for it. Idempotent with the EnvironmentRequested
+        // path above: whichever runs first configures the view and the other finds it already done.
+        Dispatcher.UIThread.Post(
+            () =>
+            {
+                AdoptRenderer();
+                input.Acquire();
+            },
+            DispatcherPriority.Loaded);
+    }
+
+    private void OnTerminalFocusRequested(object? sender, EventArgs e) =>
+        Dispatcher.UIThread.Post(input.Acquire, DispatcherPriority.Loaded);
+
+    /// <summary>
+    /// Takes the keyboard off the renderer and puts the taps on it to use.
+    /// </summary>
+    /// <remarks>
+    /// See <see cref="RendererFocus"/> for why a <c>WebView</c> that cannot hold focus still scrolls, and why
+    /// the tap has to be reported rather than left to the page.
+    /// </remarks>
+    private void AdoptRenderer() => RendererFocus.Suppress(input.Acquire);
+
     /// <summary>
     /// Puts the keyboard away on the way to the connect menu.
     /// </summary>
@@ -137,50 +259,28 @@ internal sealed partial class TerminalScreen : UserControl
 
     private void BuildAccessoryRow()
     {
-        var row = this.FindControl<StackPanel>("AccessoryKeys")!;
+        // ── nothing to hand back any more ──────────────────────────────────────────────────────────────
+        // This row used to counter both halves of every press with a posted TerminalFocus.Return(), because
+        // the platform requests Android's own focus for Avalonia's view after dispatching each handled
+        // touch, and the keyboard belonged to the WebView's input connection. Under ADR 0015 it belongs to
+        // TerminalInputSurface, which is an Avalonia control — so the focus the platform takes is the focus
+        // the keyboard already wants, and the counter-move is not merely unnecessary but wrong. The keys
+        // stay Focusable=false, which is now what keeps Avalonia's own focus on the input surface.
 
-        foreach (var (label, bytes, latches) in Keys)
+        Fill(this.FindControl<UniformGrid>("AccessoryKeys")!, ModifierKeys);
+        Fill(this.FindControl<UniformGrid>("ArrowKeys")!, CursorKeys);
+    }
+
+    /// <summary>Puts one table's worth of keys into one row of the keypad.</summary>
+    /// <remarks>
+    /// The rows differ in what they hold and in nothing else — same control, same class, same wiring — so
+    /// they are filled by one method rather than by two that would drift.
+    /// </remarks>
+    private void Fill(UniformGrid row, (string Label, byte[] Bytes, bool Latches)[] keys)
+    {
+        foreach (var (label, bytes, latches) in keys)
         {
-            var key = new Button
-            {
-                Content = new TextBlock
-                {
-                    Text = label,
-                    FontFamily = (FontFamily)Application.Current!.FindResource("MonoFont")!,
-                    FontSize = 11,
-                    HorizontalAlignment = HorizontalAlignment.Center,
-                    VerticalAlignment = VerticalAlignment.Center,
-                },
-
-                // 44 wide, and that is the number that matters: the design draws these flexed across the
-                // width, which at 360dp with ten keys is 32 pixels each — under every thumb-target
-                // guideline there is. The height came down with the row it sits in, from 38 to 30, and it
-                // costs nothing a width does: the keys are a single row with the terminal above and the
-                // system's gesture bar below, so there is no neighbour a short press can land on instead.
-                MinWidth = 44,
-                Height = 30,
-                Padding = new Thickness(10, 0),
-                CornerRadius = new CornerRadius(9),
-                Background = Palette("Panel"),
-                BorderBrush = Palette("BorderMid"),
-                BorderThickness = new Thickness(1),
-                Foreground = Palette("TextDim"),
-                HorizontalContentAlignment = HorizontalAlignment.Center,
-
-                // ◆ NOT FOCUSABLE, AND THAT IS THE WHOLE CONTROL RATHER THAN A DETAIL.
-                //
-                // These keys are an extension of the keyboard, not a place the keyboard should go. As
-                // ordinary buttons they took Avalonia's focus on tap, which takes it off the NativeWebView
-                // — and the package's own OnLostFocus then calls the adapter's ResignFocus(). So pressing
-                // Tab or an arrow handed the terminal one byte and took the keyboard away from it: the next
-                // thing typed on a hardware keyboard went nowhere, and the row went on working because its
-                // buttons are pressed rather than typed into, which is what makes it look like the terminal
-                // had died instead.
-                //
-                // Focusable=false is what a toolbar button is, and it means the focused element never
-                // changes: the WebView is still it, so nothing resigns and nothing has to be handed back.
-                Focusable = false,
-            };
+            var key = CreateKey(label);
 
             if (latches)
             {
@@ -196,6 +296,39 @@ internal sealed partial class TerminalScreen : UserControl
         }
     }
 
+    /// <summary>One key of the accessory row, before its click is wired.</summary>
+    /// <remarks>Split from <see cref="BuildAccessoryRow"/> for length rather than for reuse.</remarks>
+    private static Button CreateKey(string label) =>
+        new()
+        {
+            Content = label,
+
+            // ◆ Every dimension and brush this used to set by hand is now Button.hotkey in Phone.axaml,
+            // and the move is what lets the row have a pressed state at all: a style can say :pressed and
+            // an assignment cannot. It also takes this file out of the business of naming the palette's
+            // brushes, which is how a keypad ends up with a grey nothing else uses.
+            //
+            // Content is the string rather than a TextBlock built here for the same reason — the class
+            // carries the font and the alignment, so a wrapper element would only be somewhere for them to
+            // disagree.
+            Classes = { "hotkey" },
+
+            // ◆ NOT FOCUSABLE, AND THAT IS THE WHOLE CONTROL RATHER THAN A DETAIL.
+            //
+            // These keys are an extension of the keyboard, not a place the keyboard should go. As
+            // ordinary buttons they take Avalonia's focus on tap, which takes it off whatever the
+            // terminal's input target is — and a modifier row that dismisses the keyboard every time it
+            // is used is not a modifier row. Focusable=false is what a toolbar button is, and it means
+            // the focused element never changes: TerminalInputSurface is still it, so the input
+            // connection is never torn down and nothing has to be handed back.
+            //
+            // This used to be only half a fix, because the target was the NativeWebView and Android keeps
+            // a focus of its own that the platform hands to Avalonia's view after every touch it
+            // dispatches. That half is gone rather than countered: under ADR 0015 the input target is an
+            // Avalonia control, so the focus Android moves is already where the keyboard should be.
+            Focusable = false,
+        };
+
     private void ToggleControl()
     {
         controlLatched = !controlLatched;
@@ -207,18 +340,12 @@ internal sealed partial class TerminalScreen : UserControl
 
         // Latched state has to be visible. A modifier that is on and does not look on is how somebody sends
         // ^L to a database prompt believing they typed an l.
-        controlKey.Background = Palette(controlLatched ? "Active" : "Panel");
-        controlKey.Foreground = Palette(controlLatched ? "AccentText" : "TextDim");
+        //
+        // A class rather than two brush assignments, so what "latched" looks like is stated once in
+        // Phone.axaml beside what "pressed" looks like — which is where somebody changing either will be.
+        controlKey.Classes.Set("latched", controlLatched);
     }
 
-    /// <summary>One brush from <c>Theme/Palette.axaml</c>, by key.</summary>
-    /// <remarks>
-    /// The accessory row is built in code because its keys come from a table, so its colours cannot be set in
-    /// XAML with the rest of the screen's. Resolving them by name is the next best thing: a row that named
-    /// its own blues is how the palette ends up with a fifth surface nobody meant to add.
-    /// </remarks>
-    private static IBrush? Palette(string key) => Application.Current?.FindResource(key) as IBrush;
-
     /// <remarks>
     /// The control translation is the ASCII one and nothing cleverer: Ctrl-A through Ctrl-Z are 0x01 to
     /// 0x1A, which is letter &amp; 0x1F. Applied only to letters, because Ctrl with an arrow key is a

src/DodoSSH.Client.Android/Views/VaultsScreen.axaml

@@ -220,7 +220,7 @@
                       <TextBlock Classes="mono" FontSize="13.5" FontWeight="SemiBold" Text="{Binding Name}"
                                  TextTrimming="CharacterEllipsis" />
                       <TextBlock Classes="detail" FontSize="10.5" Text="{Binding Email}"
-                                 TextTrimming="CharacterEllipsis" />
+                                 TextWrapping="NoWrap" TextTrimming="CharacterEllipsis" />
 
                       <!--
                         ◆ The one fact on this row that decides whether the button at the foot of the screen
@@ -336,9 +336,11 @@
       the thing the whole screen is about — a member can only be selected under one, so choosing who is the
       only half left to make.
     -->
+    <!-- DeepChrome rather than Chrome, since v5 — the same raised-bar decision SnippetsScreen and
+         FilesScreen make for their own selection bars; see the remark on PhoneShell's vault header. -->
     <Border Grid.Row="4"
             IsVisible="{Binding SelectedMember, Converter={x:Static ObjectConverters.IsNotNull}}"
-            Background="{StaticResource Chrome}" BorderBrush="{StaticResource Border}"
+            Background="{StaticResource DeepChrome}" BorderBrush="{StaticResource Border}"
             BorderThickness="0,1,0,0" Padding="14,12">
       <StackPanel Spacing="9">
 

src/DodoSSH.Client.Android/packages.lock.json

@@ -74,9 +74,9 @@
       },
       "Microsoft.NET.ILLink.Tasks": {
         "type": "Direct",
-        "requested": "[10.0.10, )",
-        "resolved": "10.0.10",
-        "contentHash": "f5VCIE7AJpd5YvzNTeMGVzQIgyE9tX+AreTYwQF+REbu+DZo/2Ae+jNSwhPEYrVz6RRkd7y8ubXjk6Nn6Ka+Cg=="
+        "requested": "[10.0.11, )",
+        "resolved": "10.0.11",
+        "contentHash": "IBf7lbovvjGWVWXZX5cJ/cO0WXbId0Zq4BuSeT94mGZuOAP66oMeH9PTBZ9Jpp3Jb6jtK0qm/NyUbPRo1gC/wQ=="
       },
       "MinVer": {
         "type": "Direct",
@@ -984,7 +984,7 @@
         "type": "Project",
         "dependencies": {
           "NSec.Cryptography": "[26.4.0, )",
-          "SSH.NET": "[2025.1.0, )"
+          "SSH.NET": "[2026.0.0, )"
         }
       },
       "dodossh.client.storage": {
@@ -1044,9 +1044,9 @@
       },
       "BouncyCastle.Cryptography": {
         "type": "CentralTransitive",
-        "requested": "[2.6.2, )",
-        "resolved": "2.6.2",
-        "contentHash": "7oWOcvnntmMKNzDLsdxAYqApt+AjpRpP2CShjMfIa3umZ42UQMvH0tl1qAliYPNYO6vTdcGMqnRrCPmsfzTI1w=="
+        "requested": "[2.7.0, )",
+        "resolved": "2.7.0",
+        "contentHash": "U+12df8UEWHgBi04YVf/Lgi2dy3SItlIYvHjjEVa/BngCQIzDCDRBk50DDByCfDvSbe5pRNFr3b7UrVK2kMcLw=="
       },
       "EFCore.NamingConventions": {
         "type": "CentralTransitive",
@@ -1140,11 +1140,11 @@
       },
       "SSH.NET": {
         "type": "CentralTransitive",
-        "requested": "[2025.1.0, )",
-        "resolved": "2025.1.0",
-        "contentHash": "jrnbtf0ItVaXAe6jE8X/kSLa6uC+0C+7W1vepcnRQB/rD88qy4IxG7Lf1FIbWmkoc4iVXv0pKrz+Wc6J4ngmHw==",
+        "requested": "[2026.0.0, )",
+        "resolved": "2026.0.0",
+        "contentHash": "Yu9dirPq8l3oaat0+OQ7K0nUf5MmYltpia5UGqsApTG4zTPvBC1cxbNnC3NERij26dUST0A3Ef1QdHSn5ArbWQ==",
         "dependencies": {
-          "BouncyCastle.Cryptography": "2.6.2",
+          "BouncyCastle.Cryptography": "2.7.0",
           "Microsoft.Extensions.Logging.Abstractions": "8.0.3"
         }
       }

src/DodoSSH.Client.Api/DodoSshApiClient.cs

@@ -259,6 +259,22 @@ public interface ISyncApi
         CancellationToken cancellationToken);
 }
 
+/// <summary>Asks the server for leave to relay one connection to one host.</summary>
+/// <remarks>
+/// The request names a host and nothing else. Where it is, and whether this caller may reach it, is the
+/// server's decision, made from its own copy of the host's relay address — which is what keeps the relay
+/// from being an open proxy that a client points wherever it likes. See ADR 0004.
+/// </remarks>
+public interface IRelayApi
+{
+    /// <summary>Issues a single-use ticket, good for thirty seconds, for one relay socket to this host.</summary>
+    /// <exception cref="DodoSshApiException">
+    /// The server refused: <c>relay-unavailable</c>, <c>relay-target-rejected</c> or
+    /// <c>relay-limit-reached</c>, each with its own remedy.
+    /// </exception>
+    Task<RelayTicketResponse> IssueRelayTicketAsync(Guid hostId, CancellationToken cancellationToken);
+}
+
 /// <summary>
 /// The typed client for one DodoSSH server.
 /// </summary>
@@ -275,7 +291,7 @@ public interface ISyncApi
 /// </para>
 /// </remarks>
 public sealed class DodoSshApiClient(HttpClient http, IAccessTokenProvider tokens)
-    : IAccountApi, ISyncApi, ITeamApi, IDirectoryApi, IVaultGrantApi
+    : IAccountApi, ISyncApi, ITeamApi, IDirectoryApi, IVaultGrantApi, IRelayApi
 {
     private const string MetaPath = "/api/v1/meta";
     private const string ConfigurationPath = "/.well-known/dodossh-configuration";
@@ -310,6 +326,20 @@ public sealed class DodoSshApiClient(HttpClient http, IAccessTokenProvider token
             DodoSshJsonContext.Default.DodoSshConfiguration,
             cancellationToken);
 
+    /// <inheritdoc />
+    /// <remarks>
+    /// Never cached, and never fetched ahead of need: a ticket is spent by the socket it opens and is
+    /// worthless thirty seconds after issue, so the only right moment to ask for one is the connect that
+    /// will use it.
+    /// </remarks>
+    public Task<RelayTicketResponse> IssueRelayTicketAsync(Guid hostId, CancellationToken cancellationToken) =>
+        SendAsync(
+            HttpMethod.Post,
+            RelayProtocol.TicketsPath,
+            JsonContent.Create(new RelayTicketRequest(hostId, PortForwardId: null), DodoSshJsonContext.Default.RelayTicketRequest),
+            DodoSshJsonContext.Default.RelayTicketResponse,
+            cancellationToken);
+
     /// <summary>
     /// Reads the caller's profile, unlock material and reachable vaults.
     /// </summary>

src/DodoSSH.Client.App/App.axaml.cs

@@ -3,6 +3,7 @@ using Avalonia.Controls;
 using Avalonia.Controls.ApplicationLifetimes;
 using Avalonia.Input.Platform;
 using Avalonia.Markup.Xaml;
+using Avalonia.Threading;
 using DodoSSH.Client.App.Platform;
 using DodoSSH.Client.App.Views;
 using DodoSSH.Client.Auth;
@@ -33,12 +34,77 @@ internal sealed partial class DodoSshApp : Application
     {
         if (ApplicationLifetime is IClassicDesktopStyleApplicationLifetime desktop)
         {
-            Compose(desktop);
+            ShowSomethingWhileComposing(desktop);
         }
 
         base.OnFrameworkInitializationCompleted();
     }
 
+    /// <summary>
+    /// Puts a window on screen first, and composes the application behind it.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// <b>The splash <em>is</em> the main window for a moment, which is what makes this work at all.</b> The
+    /// desktop lifetime shows whatever <c>MainWindow</c> holds once this method returns and then runs the
+    /// message loop; a window created here and merely shown would be a window with no lifetime attached, and
+    /// with <c>ShutdownMode.OnLastWindowClose</c> the first close of anything would end the process. Handing
+    /// the property over is one assignment, and Avalonia reads it lazily.
+    /// </para>
+    /// <para>
+    /// <b>And composition is posted rather than called.</b> It is synchronous and takes long enough to be
+    /// worth a splash in the first place — which is the whole argument for this method — so running it here
+    /// would block the very dispatcher turn that has to paint the thing. Background priority puts it behind
+    /// rendering, which is what makes this a painted window rather than merely a shown one.
+    /// </para>
+    /// </remarks>
+    private static void ShowSomethingWhileComposing(IClassicDesktopStyleApplicationLifetime desktop)
+    {
+        var splash = new SplashWindow();
+
+        desktop.MainWindow = splash;
+
+        // Fully qualified, and it has to be: Application inherits an instance Dispatcher property from
+        // AvaloniaObject, member lookup beats the namespace, and a static method here cannot reach an
+        // instance property — so the unqualified name is a compile error rather than the type. Same below.
+        Avalonia.Threading.Dispatcher.UIThread.Post(
+            () => ComposeBehind(desktop, splash),
+            DispatcherPriority.Background);
+    }
+
+    /// <remarks>
+    /// <para>
+    /// The order at the end is not interchangeable. <c>ShutdownMode</c> is <c>OnLastWindowClose</c>, so the
+    /// real window is shown while the splash is still open and the splash is closed afterwards; closing the
+    /// only window there is would end the process instead of handing it over.
+    /// </para>
+    /// <para>
+    /// <b>A failure is caught rather than left to the dispatcher, and the splash is exactly why.</b> Before
+    /// this method existed a composition that threw took the process with it and left nothing behind. Now
+    /// there is something to leave behind — so an exception escaping a posted callback would abandon a window
+    /// saying "Starting…" over a process that had stopped, which is worse than the crash it replaced. What
+    /// arrives instead is the sentence and a way to close it.
+    /// </para>
+    /// </remarks>
+    private static void ComposeBehind(
+        IClassicDesktopStyleApplicationLifetime desktop,
+        SplashWindow splash)
+    {
+        try
+        {
+            Compose(desktop);
+        }
+        catch (Exception exception) when (exception is not OutOfMemoryException)
+        {
+            splash.CouldNotStart(exception.Message);
+
+            return;
+        }
+
+        desktop.MainWindow?.Show();
+        splash.Close();
+    }
+
     /// <remarks>
     /// <para>
     /// Composed by hand rather than through a container. The graph is a handful of objects deep and an
@@ -72,6 +138,17 @@ internal sealed partial class DodoSshApp : Application
             }
         };
 
+    /// <summary>Reads this machine's clipboard.</summary>
+    /// <remarks>
+    /// The other direction, looked up per call for the same reason and with the same fall-through. One
+    /// caller: the terminal's right-click paste, which the renderer cannot serve itself — see
+    /// <c>TerminalClientOpcode.ClipboardPaste</c>.
+    /// </remarks>
+    private static Func<Task<string?>> ClipboardReader(IClassicDesktopStyleApplicationLifetime desktop) =>
+        async () => TopLevel.GetTopLevel(desktop.MainWindow) is { Clipboard: { } clipboard }
+            ? await clipboard.TryGetTextAsync().ConfigureAwait(false)
+            : null;
+
     /// <summary>The terminal workspace, with its loopback listener already up.</summary>
     /// <remarks>
     /// Extracted so that constructing it and starting it cannot drift apart: the data plane's socket has to
@@ -92,7 +169,10 @@ internal sealed partial class DodoSshApp : Application
 
     private static void Compose(IClassicDesktopStyleApplicationLifetime desktop)
     {
-        var paths = ClientPaths.Default;
+        // ForChannel rather than Default, so a nightly keeps its cache, outbox and device key somewhere
+        // the release build never opens. See ClientPaths.ForChannel for why sharing them is the failure
+        // worth spending a directory on.
+        var paths = ClientPaths.ForChannel(DesktopChannel.Name);
         var caches = ClientCacheFactory.ForFile(paths.CacheFile);
 
         // Known hosts live in the vault, so trust survives a restart and follows the user to every device.
@@ -110,14 +190,75 @@ internal sealed partial class DodoSshApp : Application
         var browser = new SystemBrowserLauncher();
 
         // Both chosen once, here, because each is a property of the machine rather than of any session. A
-        // computer with a usable TPM gets the store that keeps a device key behind a Windows consent prompt;
-        // anything else gets one that reports itself unavailable, so unlock keeps asking for the passphrase
+        // Windows box with a usable TPM gets a consent prompt, a Mac with a Secure Enclave gets Touch ID, a
+        // Linux box whose TPM this user may open gets a PIN the TPM checks — see PinPrompt — and anything
+        // else gets a store that reports itself unavailable, so unlock keeps asking for the passphrase
         // (ADR 0007). The update channel answers the same shape of question about how this copy was
         // installed, and a build run from a checkout likewise gets one that says so. See ADR 0013.
-        var deviceKeys = DesktopDeviceKeyStores.ForThisMachine(paths);
+        var deviceKeys = DesktopDeviceKeyStores.ForThisMachine(paths, PinPrompt(desktop));
         var updates = UpdateChannels.ForThisMachine();
 
-        var viewModel = new MainWindowViewModel(
+        var viewModel = BuildShell(desktop, paths, caches, workspace, knownHosts, deviceKeys, connections,
+            browser, updates);
+
+        desktop.MainWindow = new MainWindow { DataContext = viewModel };
+
+        // The other half of Program's claim on this profile: a second launch was refused, and this is what
+        // answers it. Posted onto the UI thread because a pipe accept completes on a thread pool thread and
+        // a window may only be touched from the thread that owns it — see SingleInstance.Listen.
+        var others = SingleInstance.Listen(
+            paths,
+            () => Avalonia.Threading.Dispatcher.UIThread.Post(() => BringToFront(desktop.MainWindow)));
+
+        // Started rather than awaited: the framework's initialisation must not block on a schema
+        // migration. The view model shows its own progress and handles its own failures, which is why
+        // discarding the task here is safe rather than merely convenient.
+        _ = viewModel.StartAsync(CancellationToken.None);
+
+        WireShutdown(desktop, viewModel, workspace, caches, others);
+    }
+
+    /// <summary>
+    /// Raises the window somebody has just asked for by launching the application again.
+    /// </summary>
+    /// <remarks>
+    /// Un-minimised first, because activating a minimised window leaves it minimised and the launch would
+    /// look ignored. <c>Activate</c> is a request rather than a guarantee — Windows refuses the foreground to
+    /// a process that did not receive the input, and what happens then is a flashing taskbar button, which is
+    /// still an answer. That is the platform's own rule about focus stealing and not a thing to work around.
+    /// </remarks>
+    private static void BringToFront(Window? window)
+    {
+        if (window is null)
+        {
+            return;
+        }
+
+        if (window.WindowState is WindowState.Minimized)
+        {
+            window.WindowState = WindowState.Normal;
+        }
+
+        window.Show();
+        window.Activate();
+    }
+
+    /// <summary>Composes the shell itself, with everything this platform answers for it.</summary>
+    /// <remarks>
+    /// Split out of <see cref="Compose"/> for length rather than for reuse — it has one caller, and the
+    /// method it came out of is at the limit the analyzers set.
+    /// </remarks>
+    private static MainWindowViewModel BuildShell(
+        IClassicDesktopStyleApplicationLifetime desktop,
+        ClientPaths paths,
+        ClientCacheFactory caches,
+        TerminalWorkspace workspace,
+        VaultKnownHostStore knownHosts,
+        IDeviceKeyStore deviceKeys,
+        SshNetConnectionFactory connections,
+        SystemBrowserLauncher browser,
+        IUpdateChannel updates) =>
+        new(
             paths,
             caches,
             workspace,
@@ -137,18 +278,88 @@ internal sealed partial class DodoSshApp : Application
                 .ConfigureAwait(false),
 
             copyToClipboard: ClipboardWriter(desktop),
-            updates: updates);
+            readClipboard: ClipboardReader(desktop),
+            updates: updates,
+            openInEditor: OpenInDefaultEditorAsync,
+            watchFile: WatchFile);
 
-        desktop.MainWindow = new MainWindow { DataContext = viewModel };
+    /// <summary>
+    /// Hands a local file to whatever this machine opens that kind of file with.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// One of the two halves of editing a remote file in place, and both are this platform's rather than
+    /// the shell's: what opens a <c>.conf</c> is a registry association, a desktop entry or a Launch
+    /// Services binding depending on which desktop this is, and watching a file is a platform object with
+    /// its own thread. The shell decides <em>when</em>; see <c>TransfersViewModel.EditRemoteAsync</c>.
+    /// </para>
+    /// <para>
+    /// <c>UseShellExecute</c> is the whole of it, and it is what makes the same three lines work on all
+    /// three desktops: Windows resolves the association, macOS goes through <c>open</c> and Linux through
+    /// <c>xdg-open</c>, each of which .NET reaches for when the flag is set. Without it the path is treated
+    /// as an executable to run, which for a <c>.conf</c> fails and for something else would be worse.
+    /// </para>
+    /// <para>
+    /// The process is not waited on and is deliberately not kept: an editor may hand off to an already-open
+    /// instance and exit at once, so its lifetime says nothing about whether the file is still being
+    /// edited. What ends an edit session is the user saying so — see the file browser's STOP EDITING.
+    /// </para>
+    /// </remarks>
+    private static Task OpenInDefaultEditorAsync(string path)
+    {
+        using var opened = System.Diagnostics.Process.Start(
+            new System.Diagnostics.ProcessStartInfo(path) { UseShellExecute = true });
 
-        // Started rather than awaited: the framework's initialisation must not block on a schema
-        // migration. The view model shows its own progress and handles its own failures, which is why
-        // discarding the task here is safe rather than merely convenient.
-        _ = viewModel.StartAsync(CancellationToken.None);
-
-        WireShutdown(desktop, viewModel, workspace, caches);
+        return Task.CompletedTask;
     }
 
+    /// <summary>
+    /// Watches one file for saves, until the handle is disposed.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// ◆ <b>The directory is watched and the name filtered, which is not the same as watching the file.</b>
+    /// A great many editors do not write in place: they write a temporary file beside the original and
+    /// rename it over the top, which to a watcher pointed at an inode is the file disappearing. Watching the
+    /// containing directory for that name catches the rename as a <c>Created</c>, which is what makes this
+    /// work with the editors people actually use.
+    /// </para>
+    /// <para>
+    /// <b>Every kind of change is one callback, and the shell debounces nothing.</b> One save can raise two
+    /// or three events; what the shell does with each is stat the host and compare, which is a round trip
+    /// but a small one, and the alternative — a timer here — is a delay before somebody's edit reaches the
+    /// server for the sake of saving it.
+    /// </para>
+    /// </remarks>
+    private static FileSystemWatcher WatchFile(string path, Action onChanged)
+    {
+        var watcher = new FileSystemWatcher(
+            Path.GetDirectoryName(path)!, Path.GetFileName(path))
+        {
+            NotifyFilter = NotifyFilters.LastWrite | NotifyFilters.Size | NotifyFilters.FileName,
+        };
+
+        watcher.Changed += (_, _) => onChanged();
+        watcher.Created += (_, _) => onChanged();
+        watcher.Renamed += (_, _) => onChanged();
+
+        watcher.EnableRaisingEvents = true;
+
+        return watcher;
+    }
+
+    /// <summary>How the Linux device key store asks for a PIN, and the only thing that ever calls it.</summary>
+    /// <remarks>
+    /// The lifetime is captured rather than the window, so this can be composed before there is one —
+    /// composition runs first, and <c>DevicePinWindow.AskAsync</c> answers null when there is nothing to
+    /// sit over. Windows and macOS never reach it: on those platforms the consent dialog is the operating
+    /// system's and this process is not in the conversation. See ADR 0007.
+    /// </remarks>
+    private static Func<DevicePinPurpose, CancellationToken, ValueTask<string?>> PinPrompt(
+        IClassicDesktopStyleApplicationLifetime desktop) =>
+        (purpose, cancellationToken) =>
+            DevicePinWindow.AskAsync(() => desktop.MainWindow, purpose, cancellationToken);
+
     /// <remarks>
     /// Shutdown is deferred rather than blocked on. Sessions hold SSH connections and a listening socket, and
     /// blocking the UI thread on their disposal is how an application comes to take several seconds to close —
@@ -158,7 +369,8 @@ internal sealed partial class DodoSshApp : Application
         IClassicDesktopStyleApplicationLifetime desktop,
         MainWindowViewModel viewModel,
         TerminalWorkspace workspace,
-        ClientCacheFactory caches)
+        ClientCacheFactory caches,
+        IAsyncDisposable others)
     {
         var shuttingDown = false;
 
@@ -177,6 +389,11 @@ internal sealed partial class DodoSshApp : Application
             await viewModel.DisposeAsync().ConfigureAwait(true);
             await workspace.DisposeAsync().ConfigureAwait(true);
 
+            // Awaited rather than dropped, so the pipe name is free before this process is. It matters for
+            // one caller and matters a lot there: an update relaunches this application, and a successor
+            // that found the name still taken would be a second instance refusing itself.
+            await others.DisposeAsync().ConfigureAwait(true);
+
             caches.Dispose();
 
             desktop.Shutdown();

src/DodoSSH.Client.App/Assets/dodossh-icon.ps1

@@ -1,13 +1,18 @@
-# Regenerates dodossh.ico from the same geometry the Android launcher icon draws.
+# Regenerates dodossh.ico and dodossh.icns from the same geometry the Android launcher icon draws.
 #
 # The phone's mark is a vector — Resources/drawable/ic_launcher_foreground.xml — and the whole
 # reason it is a vector is that there is then one geometry to change and no set of PNG densities
-# to forget one of. Windows will not take a vector: <ApplicationIcon> wants an .ico and nothing
-# else, and Window.Icon wants a bitmap. So the raster exists, and this script is how it stays
-# honest: the numbers below are the ones in that XML, and regenerating is the whole edit.
+# to forget one of. Neither desktop platform will take a vector: <ApplicationIcon> wants an .ico
+# and nothing else, Window.Icon wants a bitmap, and vpk wants an .icns for the macOS bundle. So
+# the rasters exist, and this script is how they stay honest: the numbers below are the ones in
+# that XML, and regenerating is the whole edit.
 #
 #     pwsh -File src/DodoSSH.Client.App/Assets/dodossh-icon.ps1
 #
+# Both outputs are written every run, deliberately. Two scripts, or one script with a switch,
+# is how the two files come to be drawn from different geometry — which nobody would notice,
+# because no one person looks at a Windows taskbar and a macOS Dock on the same afternoon.
+#
 # Coordinates are the launcher's 108-unit viewport, mapped so the middle 72 fills the canvas.
 # That 72 is not an arbitrary crop: it is the part of an adaptive icon a launcher actually shows,
 # the outer 18 on each edge being what it eats for masking and parallax. Rendering the whole 108
@@ -22,37 +27,57 @@ Set-StrictMode -Version Latest
 $ErrorActionPreference = 'Stop'
 Add-Type -AssemblyName System.Drawing
 
-$accent = [System.Drawing.ColorTranslator]::FromHtml('#5B8CFF')  # dodo_accent / AccentColor
-$ink = [System.Drawing.ColorTranslator]::FromHtml('#0E1220')     # AccentInk
+$accent = [System.Drawing.ColorTranslator]::FromHtml('#5D42DE')  # dodo_accent / AccentColor
+$ink = [System.Drawing.ColorTranslator]::FromHtml('#FFFFFF')     # AccentInk
 
 # Every size Windows asks for: 16 in a titlebar and a tree, 32 on the desktop, 48 in a large-icon
 # view, 256 for the preview pane. Shipping fewer means Windows downsamples one of these to fill
 # the gap, and its downsampler is not kind to a hairline.
 $sizes = @(16, 20, 24, 32, 40, 48, 64, 128, 256)
 
-function New-MarkPng([int]$size)
+# $tileFraction is how much of the canvas the accent tile fills, and it is the one number that
+# differs between the two platforms.
+#
+# Windows passes 1.0: the tile bleeds to the edge, because Windows draws application icons at
+# whatever size they come in and every other icon on the taskbar does the same.
+#
+# macOS passes 0.8047, and that is not taste. Apple's icon grid puts a rounded-rect app icon in
+# an 824-pixel square inside a 1024-pixel canvas — 824/1024 — with the remaining hundred pixels a
+# side left as air for the Dock's shadow and its magnification. An icon that ignores the grid and
+# bleeds to the edge does not read as bold; it reads as the one icon in the Dock that is too big,
+# because it sits beside Finder and Safari which do not.
+function New-MarkPng([int]$size, [double]$tileFraction = 1.0)
 {
     $bitmap = New-Object System.Drawing.Bitmap($size, $size, [System.Drawing.Imaging.PixelFormat]::Format32bppArgb)
     $g = [System.Drawing.Graphics]::FromImage($bitmap)
     $g.SmoothingMode = [System.Drawing.Drawing2D.SmoothingMode]::AntiAlias
     $g.PixelOffsetMode = [System.Drawing.Drawing2D.PixelOffsetMode]::HighQuality
 
+    # The tile, and the inset that centres it when it does not fill the canvas.
+    $tile = [double]$size * $tileFraction
+    $inset = ([double]$size - $tile) / 2.0
+
     # The accent tile, rounded as a launcher mask rounds it. A square-cornered tile would be the
     # one icon on the taskbar with corners, which reads as unfinished rather than as deliberate.
-    $radius = [double]$size * 0.22
+    #
+    # 0.22 of the tile rather than of the canvas, so the corner keeps its proportion to the shape
+    # it is rounding instead of growing as the air around it does. It is also within a whisker of
+    # the 185/824 Apple's own grid specifies, which is why one radius serves both files.
+    $radius = $tile * 0.22
     $d = $radius * 2.0
     $path = New-Object System.Drawing.Drawing2D.GraphicsPath
-    $path.AddArc(0.0, 0.0, $d, $d, 180, 90)
-    $path.AddArc($size - $d, 0.0, $d, $d, 270, 90)
-    $path.AddArc($size - $d, $size - $d, $d, $d, 0, 90)
-    $path.AddArc(0.0, $size - $d, $d, $d, 90, 90)
+    $path.AddArc($inset, $inset, $d, $d, 180, 90)
+    $path.AddArc($inset + $tile - $d, $inset, $d, $d, 270, 90)
+    $path.AddArc($inset + $tile - $d, $inset + $tile - $d, $d, $d, 0, 90)
+    $path.AddArc($inset, $inset + $tile - $d, $d, $d, 90, 90)
     $path.CloseFigure()
     $brush = New-Object System.Drawing.SolidBrush($accent)
     $g.FillPath($brush, $path)
 
-    # 108-viewport units to pixels, with the outer 18 dropped on each edge.
-    $scale = [double]$size / 72.0
-    function P([double]$x, [double]$y) { New-Object System.Drawing.PointF((($x - 18.0) * $scale), (($y - 18.0) * $scale)) }
+    # 108-viewport units to pixels, with the outer 18 dropped on each edge. Scaled to the tile and
+    # offset by the inset, so the glyph keeps its place within the tile at either fraction.
+    $scale = $tile / 72.0
+    function P([double]$x, [double]$y) { New-Object System.Drawing.PointF((($x - 18.0) * $scale + $inset), (($y - 18.0) * $scale + $inset)) }
 
     # A stroke thinner than a pixel renders as a grey suggestion of itself, which at 16px is the
     # difference between a mark and a smudge. The phone's file already bumps this width for the
@@ -117,3 +142,120 @@ $target = Join-Path $PSScriptRoot 'dodossh.ico'
 $w.Dispose(); $out.Dispose()
 
 Write-Output "Wrote $target ($($sizes.Count) sizes, $((Get-Item $target).Length) bytes)"
+
+# ---- dodossh.icns, for the macOS bundle ----------------------------------------------------------
+#
+# Written here rather than by `iconutil` on a Mac, and that is the point of doing it the long way.
+# iconutil is the documented tool and it exists only on macOS, so an icon that needed it could not
+# be regenerated on the machine this project is developed on — the geometry above would change and
+# the .icns would quietly keep the old mark until somebody next opened a Mac. The container format
+# is a magic word, a length and a run of typed PNG chunks, which is little enough to own.
+#
+# ◆ EVERY LENGTH IN THIS FILE IS BIG-ENDIAN, AND BinaryWriter IS NOT.
+#
+# The one thing that will catch anybody editing this. A .icns written little-endian is not rejected
+# with an error — Finder and vpk both just show the placeholder icon, because the first chunk claims
+# a length of about two billion and the parser walks off the end and gives up. Hence Write-BE32.
+#
+# Type codes are Apple's, and the pairs are not redundant. ic08 and ic13 are both 256 pixels because
+# one is "256 at 1x" and the other is "128 at 2x", and a Retina display asked for the second will not
+# accept the first. Same for ic09/ic14 at 512. iconutil emits both from an .iconset for this reason,
+# so this does too.
+$icnsTypes = @(
+    @{ Type = 'ic11'; Size = 32 }    # 16@2x
+    @{ Type = 'ic12'; Size = 64 }    # 32@2x
+    @{ Type = 'ic07'; Size = 128 }   # 128@1x
+    @{ Type = 'ic13'; Size = 256 }   # 128@2x
+    @{ Type = 'ic08'; Size = 256 }   # 256@1x
+    @{ Type = 'ic14'; Size = 512 }   # 256@2x
+    @{ Type = 'ic09'; Size = 512 }   # 512@1x
+    @{ Type = 'ic10'; Size = 1024 }  # 512@2x
+)
+
+# Apple's icon grid: an 824-pixel shape centred in a 1024-pixel canvas. See New-MarkPng.
+$macTileFraction = 824.0 / 1024.0
+
+# Rendered once per distinct pixel size rather than once per type code, so the two 256s and the two
+# 512s are byte-identical and the file does not carry the same image twice over at different
+# compression. It also halves the drawing, which at 1024 is not nothing.
+$rendered = @{}
+foreach ($size in ($icnsTypes.Size | Sort-Object -Unique))
+{
+    [byte[]]$png = New-MarkPng $size $macTileFraction
+    $rendered[$size] = $png
+}
+
+$icns = New-Object System.IO.MemoryStream
+
+function Write-BE32([System.IO.Stream]$stream, [uint32]$value)
+{
+    $bytes = [System.BitConverter]::GetBytes($value)
+    if ([System.BitConverter]::IsLittleEndian) { [array]::Reverse($bytes) }
+    $stream.Write($bytes, 0, 4)
+}
+
+function Write-Ascii([System.IO.Stream]$stream, [string]$text)
+{
+    $bytes = [System.Text.Encoding]::ASCII.GetBytes($text)
+    $stream.Write($bytes, 0, $bytes.Length)
+}
+
+# The header's length field covers the whole file including the header, so it is written last —
+# eight bytes of nothing now, seeked back to and filled in once the total is known.
+Write-Ascii $icns 'icns'
+Write-BE32 $icns 0
+
+foreach ($entry in $icnsTypes)
+{
+    $payload = $rendered[$entry.Size]
+    Write-Ascii $icns $entry.Type
+
+    # Length includes this chunk's own eight-byte header, which is the off-by-eight everybody
+    # writes once.
+    Write-BE32 $icns ([uint32]($payload.Length + 8))
+    $icns.Write($payload, 0, $payload.Length)
+}
+
+$total = [uint32]$icns.Length
+$icns.Position = 4
+Write-BE32 $icns $total
+
+$icnsTarget = Join-Path $PSScriptRoot 'dodossh.icns'
+[System.IO.File]::WriteAllBytes($icnsTarget, $icns.ToArray())
+$icns.Dispose()
+
+Write-Output "Wrote $icnsTarget ($($icnsTypes.Count) entries, $((Get-Item $icnsTarget).Length) bytes)"
+
+# ---- dodossh-<size>.png, for the Linux desktop entry ---------------------------------------------
+#
+# The third platform, and the only one that wants the raster loose rather than in a container. A
+# freedesktop icon theme is a directory tree — hicolor/48x48/apps/dodossh.png and friends — and the
+# RPM built by scripts/ci-linux-rpm.sh installs one file per size into it.
+#
+# Written here for the reason this whole file exists: one geometry, regenerated in one command. The
+# alternative was to have the packaging script pull the 256px frame out of dodossh.ico at build time,
+# which works — the .ico's entries really are PNGs — and puts a second place that has to know how an
+# ICO directory is laid out. It also silently ties the Linux icon to whatever sizes Windows happens
+# to want.
+#
+# ◆ THE TILE BLEEDS TO THE EDGE HERE, AS IT DOES ON WINDOWS.
+#
+# tileFraction stays 1.0. The 0.8047 inset above is Apple's icon grid and nothing else — GNOME, KDE
+# and every panel that reads hicolor draw the icon at the size they asked for, with no shadow and no
+# magnification to leave air for. An inset icon would simply be the small one in the launcher.
+#
+# ◆ AND THE SIZES ARE NOT A ROUND NUMBER OF THEM.
+#
+# 48 is the size a GNOME dash and most panels ask for, 128 and 256 are what a large-icon file manager
+# and GNOME Software use, and 64 is between them because the 48 upscaled to it is visibly soft. There
+# is deliberately nothing below 48: the hairline floor in New-MarkPng keeps a 16px mark legible, but
+# no Linux desktop asks this theme for one — window titlebars there draw no icon at all.
+$linuxSizes = @(48, 64, 128, 256)
+
+foreach ($size in $linuxSizes)
+{
+    [byte[]]$png = New-MarkPng $size
+    $pngTarget = Join-Path $PSScriptRoot "dodossh-$size.png"
+    [System.IO.File]::WriteAllBytes($pngTarget, $png)
+    Write-Output "Wrote $pngTarget ($((Get-Item $pngTarget).Length) bytes)"
+}

src/DodoSSH.Client.App/DodoSSH.Client.App.csproj

@@ -42,6 +42,45 @@
     -->
     <ApplicationIcon>Assets/dodossh.ico</ApplicationIcon>
 
+    <!--
+      ============ WHICH CHANNEL THIS BUILD BELONGS TO ============
+
+      The same split the Android head has, for the same reason and with one difference worth stating up
+      front: Android gets separation for free, because the platform refuses an update signed by a
+      different key, so its two channels cannot replace one another whatever anybody does. Nothing
+      refuses anything here. Velopack applies what its feed serves, so the separation has to be built:
+      two pack ids, two Velopack channels, two feeds, and a profile directory each.
+
+      release — DodoSSH.Desktop, Velopack channel win, read from the newest non-prerelease release.
+                Cut from a v* tag by scripts/release-windows.ps1, by a person. See ADR 0013 rule 3.
+      nightly — DodoSSH.Desktop.Nightly, Velopack channel win-nightly, read from a prerelease release
+                CI replaces on every push to main.
+
+      Default release, so an unqualified `dotnet build` is the real application and the nightly is the
+      one you have to ask for.
+
+      What this property does *not* decide is the pack id or the title. Those are arguments to vpk and
+      live where the packaging happens — in ci.yml for the nightly and in the release script for the
+      release. Putting them here would suggest the build knows which package it will end up inside,
+      and it does not.
+    -->
+    <DodoChannel Condition="'$(DodoChannel)' == ''">release</DodoChannel>
+
+    <!--
+      For the macOS keychain interop in Platform/, and for nothing else.
+
+      Set on this project rather than in Directory.Build.props deliberately. The frameworks that hold a
+      Secure Enclave key take CFDictionaries of raw pointers, so building one means pinning arrays and
+      taking their addresses — see MacDeviceKeyStore. Every other project here is managed code with no
+      business doing that, and a solution-wide flag would quietly permit it everywhere, including in the
+      crypto project where a stray pointer is the last thing anybody wants to have been allowed.
+
+      The alternative — GCHandle.Alloc with GCHandleType.Pinned — needs no flag and was considered. It
+      would replace each `fixed` with an allocate/free pair that has to be balanced by hand across the
+      early returns those methods are full of, which trades a compiler-checked scope for a manual one.
+    -->
+    <AllowUnsafeBlocks>true</AllowUnsafeBlocks>
+
     <!--
       False here, unlike every server project. The root Directory.Build.props sets it true because
       the API is container-hosted, UTC-only and has no business formatting anything for a human.
@@ -60,6 +99,16 @@
     <AvaloniaResource Include="Assets/dodossh.ico" />
   </ItemGroup>
 
+  <ItemGroup>
+    <!--
+      Which channel this build belongs to, carried in the assembly the same way the Android head carries
+      it. Metadata rather than a compile-time constant for the reason stated there: the updater needs the
+      string rather than a branch, and a value baked into the assembly is one a crash report can be asked
+      for. See DesktopChannel, which is the only thing that reads it.
+    -->
+    <AssemblyMetadata Include="DodoChannel" Value="$(DodoChannel)" />
+  </ItemGroup>
+
   <!--
     Two native symbol files, and they are the reason a self-contained publish weighed 227 MB.
 
@@ -102,6 +151,14 @@
       interface declared in DodoSSH.Client.Session.
     -->
     <PackageReference Include="Velopack" />
+
+    <!--
+      Here for the same reason and with the same boundary as Velopack above: LinuxDeviceKeyStore is a
+      Linux-only implementation of an interface declared in DodoSSH.Client.Session, and it lives in
+      Platform/ beside the Windows and macOS stores. The Android head has its own keystore and must
+      never see this one.
+    -->
+    <PackageReference Include="Microsoft.TSS" />
   </ItemGroup>
 
   <ItemGroup>

src/DodoSSH.Client.App/Platform/DesktopChannel.cs

@@ -0,0 +1,66 @@
+using System.Reflection;
+
+namespace DodoSSH.Client.App.Platform;
+
+/// <summary>
+/// Which of the two desktop channels this build belongs to.
+/// </summary>
+/// <remarks>
+/// <para>
+/// The counterpart of the Android head's <c>DodoChannel</c> metadata, read the same way and for the same
+/// reasons. Three things depend on the answer and they are listed here rather than discovered one at a
+/// time: which Velopack channel the updater reads, whether that read considers prereleases, and which
+/// profile directory this copy keeps its cache and device key in.
+/// </para>
+/// <para>
+/// <b>Android gets this separation from the platform and this head has to build it.</b> Two Android
+/// channels cannot replace one another because the installer refuses a package signed by a different key.
+/// Nothing refuses anything on Windows: Velopack applies what its feed serves, without verifying a
+/// signature. So the two channels are kept apart by construction here — a pack id each, so they install
+/// in different directories; a Velopack channel each, so neither ever reads the other's release index; and
+/// a profile directory each, so a nightly cannot migrate the schema of a cache the release build is using.
+/// See <c>docs/adr/0013-desktop-distribution-and-updates.md</c> rule 9.
+/// </para>
+/// <para>
+/// Resolved once, at first use. The channel of a build does not change while it runs, and a value read
+/// per call site is one that eventually gets read differently in two places.
+/// </para>
+/// </remarks>
+internal static class DesktopChannel
+{
+    /// <summary>The build this channel is not: the one a person cuts from a tag.</summary>
+    internal const string Release = "release";
+
+    /// <summary>The build CI publishes from main, which installs beside the release one.</summary>
+    internal const string Nightly = "nightly";
+
+    /// <summary>
+    /// What this build says it is, defaulting to the release channel.
+    /// </summary>
+    /// <remarks>
+    /// The csproj always writes the metadata, so the fallback covers a build that reached here some other
+    /// way — an assembly loaded by a test host, a designer, a trimmed-down copy. <see cref="Release"/> is
+    /// the safe answer for all of them: it is what the value was before there were two channels, and every
+    /// consequence of being wrong about it is inert. A build that is not installed has no updater at all
+    /// (see <c>UpdateChannels.ForThisMachine</c>), and the profile directory it names is the one every
+    /// existing install already uses.
+    /// </remarks>
+    internal static string Name { get; } = Read();
+
+    /// <summary>Whether this build belongs to the nightly channel.</summary>
+    internal static bool IsNightly => string.Equals(Name, Nightly, StringComparison.Ordinal);
+
+    private static string Read()
+    {
+        var declared = typeof(DesktopChannel).Assembly
+            .GetCustomAttributes<AssemblyMetadataAttribute>()
+            .FirstOrDefault(attribute => string.Equals(attribute.Key, "DodoChannel", StringComparison.Ordinal))
+            ?.Value;
+
+        // Only the two the csproj declares are honoured. An unrecognised value is a build made by
+        // something nobody here wrote, and answering "release" to it is the same inert default as
+        // answering it to no value at all — rather than pointing an updater at a feed named by a string
+        // of unknown origin.
+        return string.Equals(declared, Nightly, StringComparison.Ordinal) ? Nightly : Release;
+    }
+}

src/DodoSSH.Client.App/Platform/LinuxDeviceKeyStore.cs

@@ -0,0 +1,491 @@
+using System.Buffers.Binary;
+using System.Runtime.Versioning;
+using System.Security.Cryptography;
+using System.Text;
+using DodoSSH.Client.Session;
+using Tpm2Lib;
+
+namespace DodoSSH.Client.App.Platform;
+
+/// <summary>Why the store is asking for a PIN.</summary>
+/// <remarks>
+/// The two cases need different words in front of a person — one is choosing a PIN and the other is
+/// entering one — and the store is the only thing that knows which is happening. Passing an enum rather
+/// than two delegates keeps the prompt a single seam.
+/// </remarks>
+public enum DevicePinPurpose
+{
+    /// <summary>A PIN is being chosen, as part of registering this machine.</summary>
+    Choosing,
+
+    /// <summary>A PIN is being asked for, to unlock.</summary>
+    Entering,
+}
+
+/// <summary>
+/// Keeps the device key sealed inside the TPM, released only against a PIN the TPM itself checks.
+/// </summary>
+/// <remarks>
+/// <para>
+/// The Linux answer to <see cref="WindowsDeviceKeyStore"/> and <c>MacDeviceKeyStore</c>, and ADR 0007's
+/// bar is the one it has to clear: <b>the enforcement must be the platform's rather than this process's.</b>
+/// A store that drew its own prompt over a file it could read anyway would be the mistake that ADR's
+/// 2026-07-30 amendment exists to record — malware running as the user skips the prompt and reads the file.
+/// </para>
+/// <para>
+/// <b>Here the TPM holds the key and the TPM checks the PIN.</b> The 32-byte device key is sealed into a
+/// keyedhash object whose <c>userAuth</c> is derived from the PIN, so unsealing is a command the TPM
+/// refuses without it. The blob on disk is encrypted to the TPM's storage hierarchy and is useless on any
+/// other machine, and the PIN is never compared by anything in this assembly. Malware running as the user
+/// can ask; it cannot answer.
+/// </para>
+/// <para>
+/// <b>And the TPM counts the failures, which is what makes a short PIN safe.</b> The sealed object
+/// deliberately does not carry <c>NoDA</c>, so every wrong answer advances the dictionary-attack counter
+/// and the TPM stops answering at all once it trips. Measured against a software TPM: four wrong PINs and
+/// it refuses the right one too, until the lockout recovers. That is the property a four-digit secret has
+/// on this platform and cannot have in a file.
+/// </para>
+/// <para>
+/// ◆ <b>What this does NOT match, stated because the difference is real and invisible.</b> Windows Hello
+/// and Touch ID raise their prompt from the platform — a secure desktop on one, a system sheet on the
+/// other — so the process asking never sees the secret. Linux has no equivalent to borrow, so the PIN is
+/// typed into a window this application draws, and a sufficiently determined attacker already running as
+/// the user could draw a convincing copy of it. The mitigation is that they could do the same for the
+/// passphrase screen, which is a much larger prize, so this is not a capability the feature creates. It is
+/// still a weaker guarantee than the other two desktops have and ADR 0007 says so rather than implying
+/// parity.
+/// </para>
+/// <para>
+/// <b>Availability still depends on reaching the device, and the package arranges that rather than asking
+/// the user to.</b> Fedora leaves <c>/dev/tpmrm0</c> as <c>root:tss</c> mode 0660, so without help this
+/// process could not open it on an ordinary desktop account. The RPM installs
+/// <c>build/linux/71-dodossh-tpm.rules</c>, which tags the device <c>uaccess</c> — logind then grants an
+/// ACL to whoever is logged in at the machine, and takes it away when they log out. Not over SSH and not
+/// to a user on an inactive VT.
+/// <para>
+/// So a build run from a checkout, or one on a machine where an administrator has masked that rule, still
+/// finds the device unreadable and reports unavailable — which keeps the passphrase path exactly as it was.
+/// The rule carries its own reasoning, including the lockout denial-of-service it accepts; see ADR 0007
+/// decision 4.
+/// </para>
+/// </remarks>
+[SupportedOSPlatform("linux")]
+public sealed class LinuxDeviceKeyStore : IDeviceKeyStore
+{
+    /// <summary>What the blob on disk starts with, so a file from elsewhere is refused rather than parsed.</summary>
+    private static readonly byte[] Magic = "DSHTPMv1"u8.ToArray();
+
+    /// <summary>
+    /// Domain separation for the PIN, so the value handed to the TPM is this application's and not the
+    /// user's PIN as typed.
+    /// </summary>
+    /// <remarks>
+    /// The TPM compares an authValue of at most one digest in length, so something has to reduce a typed
+    /// string to bytes; hashing with a label rather than truncating means a PIN reused elsewhere does not
+    /// arrive at the TPM as the same bytes it does there. It buys nothing against an attacker who has the
+    /// blob and the machine — they have this constant too — and it costs one line.
+    /// </remarks>
+    private const string PinLabel = "dodossh/devicekey/pin/v1";
+
+    private readonly ClientPaths paths;
+    private readonly Func<DevicePinPurpose, CancellationToken, ValueTask<string?>> askForPin;
+    private readonly Func<Tpm2Device> openDevice;
+
+    /// <summary>Creates a store against this machine's TPM.</summary>
+    /// <param name="paths">Where the sealed blob lives.</param>
+    /// <param name="askForPin">
+    /// Shows the PIN prompt and answers with what was typed, or null if the user declined. Declining is
+    /// ordinary — see <see cref="IDeviceKeyStore"/> — and lands on the passphrase.
+    /// </param>
+    public LinuxDeviceKeyStore(
+        ClientPaths paths,
+        Func<DevicePinPurpose, CancellationToken, ValueTask<string?>> askForPin)
+        : this(paths, askForPin, static () => new LinuxTpmDevice())
+    {
+    }
+
+    /// <summary>Creates a store against a named TPM, which is how the tests reach a simulator.</summary>
+    /// <remarks>
+    /// The seam exists because this class is testable in a way the other two stores are not, and that is
+    /// worth keeping. Windows and macOS put the prompt inside the platform, so a test that exercised
+    /// sealing would block forever on a dialog — <c>WindowsDeviceKeyStoreTests</c> records the ten minutes
+    /// that took to discover. Here the prompt is a delegate and the TPM is a device, so both can be
+    /// supplied and the whole path runs unattended.
+    /// </remarks>
+    internal LinuxDeviceKeyStore(
+        ClientPaths paths,
+        Func<DevicePinPurpose, CancellationToken, ValueTask<string?>> askForPin,
+        Func<Tpm2Device> openDevice)
+    {
+        ArgumentNullException.ThrowIfNull(paths);
+        ArgumentNullException.ThrowIfNull(askForPin);
+        ArgumentNullException.ThrowIfNull(openDevice);
+
+        this.paths = paths;
+        this.askForPin = askForPin;
+        this.openDevice = openDevice;
+    }
+
+    /// <summary>
+    /// The parent every sealed object hangs off, as a template rather than as a stored handle.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// <b>Recreated from a fixed template on every use, and not made persistent.</b> A persistent handle
+    /// would occupy one of the TPM's small number of NV slots forever, would have to be cleaned up by
+    /// <see cref="ForgetAsync"/>, and would collide with whatever else on the machine had chosen the same
+    /// index. Deriving it costs a command and leaves nothing behind.
+    /// </para>
+    /// <para>
+    /// <b>ECC P-256 rather than RSA-2048, and that is a user-visible choice.</b> The two are equally
+    /// standard as storage roots. Creating an RSA primary on a discrete TPM takes seconds — the key is
+    /// generated, not derived — and this runs on the unlock path, so it would show up as the application
+    /// hanging every time somebody opened their vault. The prototype this was built from used RSA and that
+    /// is exactly what it was changed for.
+    /// </para>
+    /// <para>
+    /// <c>NoDA</c> is set <em>here</em> and deliberately not on the sealed object. The parent has no
+    /// authValue to guess, so counting failures against it would only give an attacker a way to lock the
+    /// machine's TPM out by asking for something they were never going to get.
+    /// </para>
+    /// </remarks>
+    private static TpmPublic PrimaryTemplate => new(
+        TpmAlgId.Sha256,
+        ObjectAttr.Restricted | ObjectAttr.Decrypt | ObjectAttr.FixedTPM | ObjectAttr.FixedParent
+            | ObjectAttr.SensitiveDataOrigin | ObjectAttr.UserWithAuth | ObjectAttr.NoDA,
+        null,
+        new EccParms(
+            new SymDefObject(TpmAlgId.Aes, 128, TpmAlgId.Cfb),
+            new NullAsymScheme(),
+            EccCurve.TpmEccNistP256,
+            new NullKdfScheme()),
+        new EccPoint());
+
+    /// <summary>
+    /// The sealed object: bytes the TPM holds, released against an authValue the TPM checks.
+    /// </summary>
+    /// <remarks>
+    /// Keyedhash with a null scheme is TPM 2.0's "just keep these bytes" object. <c>UserWithAuth</c> is
+    /// what makes the authValue the thing that opens it, and the deliberate <em>absence</em> of
+    /// <c>NoDA</c> is what puts every wrong PIN on the dictionary-attack counter. Removing that one
+    /// attribute would leave a four-digit secret that could be guessed at full speed by anything holding
+    /// the blob and the machine, which is the whole of the protection this class offers.
+    /// </remarks>
+    private static TpmPublic SealedTemplate => new(
+        TpmAlgId.Sha256,
+        ObjectAttr.FixedTPM | ObjectAttr.FixedParent | ObjectAttr.UserWithAuth,
+        null,
+        new KeyedhashParms(new NullSchemeKeyedhash()),
+        new Tpm2bDigestKeyedhash());
+
+    /// <summary>Whether this machine has a TPM this process may use.</summary>
+    /// <remarks>
+    /// Probes by creating and flushing a primary rather than by looking for the device node, for the
+    /// reason <c>WindowsDeviceKeyStore.IsSupported</c> gives about its own provider: the node can be
+    /// present on a machine where the first real command then fails. Permissions are the common case here
+    /// and they fail at open; an owner hierarchy with an authValue set is the rarer one and fails at
+    /// CreatePrimary. Neither is a fault and both mean the same thing.
+    /// </remarks>
+    internal static bool IsSupported() => IsSupported(static () => new LinuxTpmDevice());
+
+    /// <inheritdoc cref="IsSupported()" />
+    internal static bool IsSupported(Func<Tpm2Device> openDevice)
+    {
+        ArgumentNullException.ThrowIfNull(openDevice);
+
+        try
+        {
+            using var device = openDevice();
+            device.Connect();
+
+            var tpm = new Tpm2(device);
+            var primary = CreatePrimary(tpm);
+
+            tpm.FlushContext(primary);
+
+            return true;
+        }
+        catch (Exception exception) when (exception is not OutOfMemoryException)
+        {
+            // Every way this fails means the same thing to the caller: no device key here, so keep asking
+            // for the passphrase. A missing device node throws IOException, a device this user may not
+            // open throws UnauthorizedAccessException, and a TPM that refuses the command throws
+            // TpmException — and on a machine with no TPM at all the constructor itself may throw before
+            // any of that.
+            return false;
+        }
+    }
+
+    /// <inheritdoc />
+    public ValueTask<bool> IsAvailableAsync(CancellationToken cancellationToken) =>
+        ValueTask.FromResult(IsSupported(openDevice));
+
+    /// <inheritdoc />
+    /// <remarks>
+    /// Asks for a PIN first and writes nothing if the user declines, which matters more here than the
+    /// symmetry suggests: <c>SaveAsync</c> is called as part of registering this machine, and a half-done
+    /// registration would leave a wrap on the server that nothing can open.
+    /// </remarks>
+    public async ValueTask SaveAsync(
+        ReadOnlyMemory<byte> devicePrivateKey,
+        CancellationToken cancellationToken)
+    {
+        var pin = await askForPin(DevicePinPurpose.Choosing, cancellationToken).ConfigureAwait(false);
+
+        if (string.IsNullOrEmpty(pin))
+        {
+            throw new OperationCanceledException("No PIN was chosen, so no device key was stored.");
+        }
+
+        var key = devicePrivateKey.ToArray();
+
+        // Off the calling thread, because every command below blocks on the chip. Creating the primary
+        // alone is hundreds of milliseconds on real hardware, and this is called from the unlock path —
+        // which on the desktop head is the UI thread.
+        var packed = await Task.Run(() => Seal(pin, key), cancellationToken).ConfigureAwait(false);
+
+        paths.EnsureCreated();
+
+        await File.WriteAllBytesAsync(paths.DeviceKeyFile, packed, cancellationToken).ConfigureAwait(false);
+    }
+
+    /// <summary>Seals the key into the TPM and answers with the blob to store.</summary>
+    private byte[] Seal(string pin, byte[] devicePrivateKey)
+    {
+        var auth = PinAuth(pin);
+
+        try
+        {
+            using var device = openDevice();
+            device.Connect();
+
+            var tpm = new Tpm2(device);
+            var primary = CreatePrimary(tpm);
+
+            try
+            {
+                var priv = tpm.Create(
+                    primary,
+                    new SensitiveCreate(auth, devicePrivateKey),
+                    SealedTemplate,
+                    null,
+                    [],
+                    out var pub,
+                    out _,
+                    out _,
+                    out _);
+
+                return Pack(pub, priv);
+            }
+            finally
+            {
+                tpm.FlushContext(primary);
+            }
+        }
+        finally
+        {
+            CryptographicOperations.ZeroMemory(auth);
+            CryptographicOperations.ZeroMemory(devicePrivateKey);
+        }
+    }
+
+    /// <inheritdoc />
+    /// <remarks>
+    /// <para>
+    /// <b>One attempt per call, and a wrong PIN is not retried here.</b> Every wrong answer advances the
+    /// TPM's lockout counter, so a loop offering three tries would spend a third of the machine's budget
+    /// on one mistyped digit. Returning null lands on the passphrase, which always works; somebody who
+    /// wants to try the PIN again asks for a device unlock again, which is a deliberate act rather than a
+    /// loop.
+    /// </para>
+    /// <para>
+    /// <b>A locked-out TPM is deliberately indistinguishable from a wrong PIN.</b>
+    /// <see cref="IDeviceKeyStore.TryLoadAsync"/> asks for exactly that and the reasoning holds here: the
+    /// caller does the same thing in every case, and "your TPM is in lockout" describes the keystore
+    /// rather than telling somebody something they can act on — the answer is still to type the
+    /// passphrase.
+    /// </para>
+    /// </remarks>
+    public async ValueTask<byte[]?> TryLoadAsync(CancellationToken cancellationToken)
+    {
+        if (!File.Exists(paths.DeviceKeyFile))
+        {
+            return null;
+        }
+
+        var stored = await File.ReadAllBytesAsync(paths.DeviceKeyFile, cancellationToken)
+            .ConfigureAwait(false);
+
+        if (!TryUnpack(stored, out var pub, out var priv))
+        {
+            return null;
+        }
+
+        // After the blob is known to be readable and before anything is asked of the user, because a
+        // prompt raised over a file this cannot use is a question with no useful answer.
+        var pin = await askForPin(DevicePinPurpose.Entering, cancellationToken).ConfigureAwait(false);
+
+        if (string.IsNullOrEmpty(pin))
+        {
+            return null;
+        }
+
+        return await Task.Run(() => Unseal(pin, pub, priv), cancellationToken).ConfigureAwait(false);
+    }
+
+    /// <summary>Asks the TPM to release the key, and answers null however it declines.</summary>
+    private byte[]? Unseal(string pin, TpmPublic pub, TpmPrivate priv)
+    {
+        var auth = PinAuth(pin);
+
+        try
+        {
+            using var device = openDevice();
+            device.Connect();
+
+            var tpm = new Tpm2(device);
+            var primary = CreatePrimary(tpm);
+
+            try
+            {
+                var loaded = tpm.Load(primary, priv, pub);
+
+                try
+                {
+                    return tpm[auth].Unseal(loaded);
+                }
+                finally
+                {
+                    tpm.FlushContext(loaded);
+                }
+            }
+            finally
+            {
+                tpm.FlushContext(primary);
+            }
+        }
+        catch (Exception exception) when (exception is not OutOfMemoryException)
+        {
+            // TpmException covers the wrong PIN and the lockout; the rest covers a TPM that has gone away
+            // between the availability check and now, which a suspend or a firmware update can do.
+            return null;
+        }
+        finally
+        {
+            CryptographicOperations.ZeroMemory(auth);
+        }
+    }
+
+    /// <inheritdoc />
+    /// <remarks>
+    /// Deleting the blob is the whole of it, and that is a property of sealing rather than an omission.
+    /// The object lives nowhere but in that file — it was never made persistent inside the TPM, see
+    /// <see cref="PrimaryTemplate"/> — so the bytes are unrecoverable the moment it is gone. There is no
+    /// TPM state to clean up and nothing left behind for a later install to trip over.
+    /// </remarks>
+    public ValueTask ForgetAsync(CancellationToken cancellationToken)
+    {
+        if (File.Exists(paths.DeviceKeyFile))
+        {
+            File.Delete(paths.DeviceKeyFile);
+        }
+
+        return ValueTask.CompletedTask;
+    }
+
+    /// <summary>Derives the TPM authValue from the PIN as typed.</summary>
+    private static byte[] PinAuth(string pin) =>
+        SHA256.HashData(Encoding.UTF8.GetBytes($"{PinLabel} {pin}"));
+
+    private static TpmHandle CreatePrimary(Tpm2 tpm) =>
+        tpm.CreatePrimary(
+            TpmRh.Owner,
+            new SensitiveCreate(),
+            PrimaryTemplate,
+            null,
+            [],
+            out _,
+            out _,
+            out _,
+            out _);
+
+    /// <summary>Puts the two halves of the sealed object in one file.</summary>
+    /// <remarks>
+    /// Both are needed to load it and neither is secret — the private half is encrypted to this TPM's
+    /// storage hierarchy, which is what makes the file useless anywhere else. Length-prefixed rather than
+    /// concatenated because their sizes vary with the TPM's algorithms, and magic-prefixed so that a file
+    /// left by another version, or by the Windows store on a dual-boot profile directory, is refused
+    /// rather than fed to the TPM as a malformed structure.
+    /// </remarks>
+    private static byte[] Pack(TpmPublic pub, TpmPrivate priv)
+    {
+        var publicPart = pub.GetTpmRepresentation();
+        var privatePart = priv.GetTpmRepresentation();
+
+        var packed = new byte[Magic.Length + 4 + publicPart.Length + 4 + privatePart.Length];
+        var at = Magic.Length;
+
+        Magic.CopyTo(packed.AsSpan());
+
+        BinaryPrimitives.WriteInt32BigEndian(packed.AsSpan(at), publicPart.Length);
+        at += 4;
+        publicPart.CopyTo(packed.AsSpan(at));
+        at += publicPart.Length;
+
+        BinaryPrimitives.WriteInt32BigEndian(packed.AsSpan(at), privatePart.Length);
+        at += 4;
+        privatePart.CopyTo(packed.AsSpan(at));
+
+        return packed;
+    }
+
+    /// <summary>Reads back what <see cref="Pack"/> wrote, or answers false.</summary>
+    private static bool TryUnpack(byte[] stored, out TpmPublic pub, out TpmPrivate priv)
+    {
+        pub = new TpmPublic();
+        priv = new TpmPrivate();
+
+        try
+        {
+            if (stored.Length < Magic.Length + 8 || !stored.AsSpan(0, Magic.Length).SequenceEqual(Magic))
+            {
+                return false;
+            }
+
+            var at = Magic.Length;
+
+            var publicLength = BinaryPrimitives.ReadInt32BigEndian(stored.AsSpan(at));
+            at += 4;
+
+            if (publicLength < 0 || at + publicLength + 4 > stored.Length)
+            {
+                return false;
+            }
+
+            var publicPart = stored.AsSpan(at, publicLength).ToArray();
+            at += publicLength;
+
+            var privateLength = BinaryPrimitives.ReadInt32BigEndian(stored.AsSpan(at));
+            at += 4;
+
+            if (privateLength < 0 || at + privateLength != stored.Length)
+            {
+                return false;
+            }
+
+            var privatePart = stored.AsSpan(at, privateLength).ToArray();
+
+            pub = Marshaller.FromTpmRepresentation<TpmPublic>(publicPart);
+            priv = Marshaller.FromTpmRepresentation<TpmPrivate>(privatePart);
+
+            return true;
+        }
+        catch (Exception exception) when (exception is not OutOfMemoryException)
+        {
+            // A blob that does not parse is a blob this machine cannot use, which is the same answer as no
+            // blob at all. Throwing here would mean a corrupted file stopped the application from starting.
+            return false;
+        }
+    }
+}

src/DodoSSH.Client.App/Platform/LinuxUpdateChannel.cs

@@ -0,0 +1,495 @@
+using System.Diagnostics;
+using System.Globalization;
+using System.Runtime.Versioning;
+using System.Security.Cryptography;
+using System.Text.Json;
+using System.Text.Json.Serialization;
+using System.Text.Json.Serialization.Metadata;
+using DodoSSH.Client.Session;
+
+namespace DodoSSH.Client.App.Platform;
+
+/// <summary>What a channel's release says it is publishing.</summary>
+/// <remarks>
+/// <para>
+/// Written by <c>scripts/ci-linux-rpm.sh</c> and read here, and it is a small file this repository
+/// controls rather than the forge's own release JSON for the reason the android manifest gives: this is
+/// fetched on a six-hourly timer, and a document whose shape changes when the forge is upgraded is not a
+/// contract.
+/// </para>
+/// <para>
+/// <b>The comparison is <see cref="RpmVersion"/> and never <see cref="Version"/>.</b> That is the string
+/// <c>rpm</c> will order the install against, so comparing anything else would let this offer an update
+/// PackageKit then refuses — the same argument <c>AndroidChannelManifest</c> makes for <c>versionCode</c>.
+/// The two differ exactly where it matters: <c>0.0.1-alpha.0.143</c> and <c>0.0.1~alpha.0.143</c> are one
+/// build, and only the second sorts below <c>0.0.1</c>.
+/// </para>
+/// </remarks>
+/// <param name="RpmVersion">The version-release string rpm knows the package by.</param>
+/// <param name="Version">What that build calls itself, for a human.</param>
+/// <param name="Rpm">The asset on the same release that holds it.</param>
+/// <param name="Sha256">The package's digest, lower-case hex.</param>
+internal sealed record LinuxChannelManifest(
+    [property: JsonPropertyName("rpmVersion")] string RpmVersion,
+    [property: JsonPropertyName("version")] string Version,
+    [property: JsonPropertyName("rpm")] string Rpm,
+    [property: JsonPropertyName("sha256")] string Sha256);
+
+/// <summary>One release as the forge describes it. Only the assets are read.</summary>
+internal sealed record LinuxForgeRelease(
+    [property: JsonPropertyName("assets")] IReadOnlyList<LinuxForgeAsset>? Assets);
+
+/// <summary>One file attached to a release.</summary>
+internal sealed record LinuxForgeAsset(
+    [property: JsonPropertyName("name")] string? Name,
+    [property: JsonPropertyName("browser_download_url")] string? DownloadUrl);
+
+[JsonSourceGenerationOptions(UnmappedMemberHandling = JsonUnmappedMemberHandling.Skip)]
+[JsonSerializable(typeof(LinuxChannelManifest))]
+[JsonSerializable(typeof(LinuxForgeRelease))]
+internal sealed partial class LinuxForgeJsonContext : JsonSerializerContext;
+
+/// <summary>
+/// Fetches a newer RPM from the project's own forge and asks the system's package manager to install it.
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>Nothing here replaces anything.</b> That is the difference from
+/// <see cref="VelopackUpdateChannel"/> and it is the decision ADR 0013 rule 11 records rather than an
+/// unfinished half: on this platform the package manager is already an updater the user trusts, audits and
+/// controls, so this class does the two things the package manager cannot — notice that a build exists and
+/// fetch it — and then hands the file to PackageKit, which raises polkit and asks. The shape is the
+/// phone's, not the desktop's: <see cref="ApplyingEndsTheProcess"/> is false, and the answer may be no.
+/// </para>
+/// <para>
+/// <b>The feed is a constant and the deployment is never asked.</b> ADR 0011 rule 2, carried unchanged
+/// from both other channels: an operator who could answer the update check could pin a chosen user to a
+/// build with a known hole, without holding any key at all.
+/// </para>
+/// <para>
+/// <b>The digest is checked and is not a signature.</b> Velopack verifies what it downloads and Android
+/// refuses a package whose signature does not match the installed one; an RPM fetched over HTTPS and
+/// handed to PackageKit has neither. So the manifest names a digest and this refuses anything else —
+/// which catches a truncated download, a caching proxy and a corrupted mirror, and catches nothing at all
+/// about who built the file, because the same job wrote both. Provenance here is the constant address; a
+/// signed repository is what would improve on it, and ADR 0013 rule 11 says why there is not one yet.
+/// </para>
+/// </remarks>
+[SupportedOSPlatform("linux")]
+internal sealed class LinuxUpdateChannel : IUpdateChannel
+{
+    /// <summary>The project's own forge, and the one address in this file.</summary>
+    /// <remarks>
+    /// The owner half of the path matters for the reason <c>AndroidUpdateChannel</c> records: the
+    /// repository moved between organisations and Gitea leaves a 301 behind, so a client naming the old
+    /// one works until somebody deletes the redirect.
+    /// </remarks>
+    private const string RepositoryApi = "https://git.dodotech.cloud/api/v1/repos/DodoTech-Public/DodoSSH";
+
+    private readonly string channel;
+    private readonly string packageName;
+    private readonly string dataDirectory;
+    private readonly HttpClient http;
+    private readonly Func<string, string?> installedVersion;
+    private readonly Action<string> handToInstaller;
+
+    private (string Version, string RpmVersion, string Url, string Sha256)? found;
+    private string? fetched;
+
+    /// <summary>Creates a channel for the running build.</summary>
+    internal LinuxUpdateChannel(string channel)
+        : this(
+            channel,
+            ClientPaths.ForChannel(DesktopChannel.Name).DataDirectory,
+            new HttpClient(),
+            InstalledRpmVersion,
+            HandToPackageKit)
+    {
+    }
+
+    /// <summary>Creates a channel with its outside world supplied, which is how the tests reach it.</summary>
+    internal LinuxUpdateChannel(
+        string channel,
+        string dataDirectory,
+        HttpClient http,
+        Func<string, string?> installedVersion,
+        Action<string> handToInstaller)
+    {
+        ArgumentException.ThrowIfNullOrEmpty(channel);
+        ArgumentException.ThrowIfNullOrEmpty(dataDirectory);
+        ArgumentNullException.ThrowIfNull(http);
+        ArgumentNullException.ThrowIfNull(installedVersion);
+        ArgumentNullException.ThrowIfNull(handToInstaller);
+
+        this.channel = channel;
+        this.dataDirectory = dataDirectory;
+        this.http = http;
+        this.installedVersion = installedVersion;
+        this.handToInstaller = handToInstaller;
+
+        packageName = string.Equals(channel, "nightly", StringComparison.Ordinal)
+            ? "dodossh-nightly"
+            : "dodossh";
+    }
+
+    /// <inheritdoc />
+    /// <remarks>
+    /// True unconditionally, because whether this copy is one that can be updated was already decided by
+    /// <see cref="UpdateChannels.ForThisMachine"/> — a build run from a checkout never reaches this class
+    /// at all. The same split Velopack's side uses, where <c>IsInstalled</c> is checked once at
+    /// composition rather than at every call site.
+    /// </remarks>
+    public bool IsSupported => true;
+
+    /// <inheritdoc />
+    public string CurrentVersion => ClientVersion.Current;
+
+    /// <inheritdoc />
+    /// <remarks>
+    /// False, and for the phone's reason rather than the desktop's. Applying hands a file to PackageKit
+    /// and returns; polkit then asks, and the user may say no. A caller that had already torn the vault
+    /// down would answer "not now" with a locked vault and every shell gone.
+    /// </remarks>
+    public bool ApplyingEndsTheProcess => false;
+
+    /// <inheritdoc />
+    /// <exception cref="HttpRequestException">The forge could not be reached, or refused.</exception>
+    /// <exception cref="InvalidOperationException">The release is there and does not carry a usable build.</exception>
+    public async Task<AvailableUpdate?> CheckAsync(CancellationToken cancellationToken)
+    {
+        var release = await ReadAsync(
+                ReleaseUrl(),
+                LinuxForgeJsonContext.Default.LinuxForgeRelease,
+                cancellationToken)
+            .ConfigureAwait(false);
+
+        var manifestName = $"linux-{channel}.json";
+
+        if (Asset(release, manifestName) is not { } manifestAsset)
+        {
+            throw new InvalidOperationException(
+                $"The {channel} release carries no {manifestName}, so there is nothing saying what it "
+                + "publishes.");
+        }
+
+        var manifest = await ReadAsync(
+                manifestAsset,
+                LinuxForgeJsonContext.Default.LinuxChannelManifest,
+                cancellationToken)
+            .ConfigureAwait(false);
+
+        if (manifest is null)
+        {
+            throw new InvalidOperationException(
+                $"The {channel} release's {manifestName} could not be read as a manifest.");
+        }
+
+        // The one place null is returned, and it means what null is documented to mean: this build is
+        // current. Compared the way rpm compares, because rpm is what accepts or refuses the install.
+        if (RpmVersion.Compare(manifest.RpmVersion, installedVersion(packageName) ?? "0") <= 0)
+        {
+            return null;
+        }
+
+        if (Asset(release, manifest.Rpm) is not { } rpm)
+        {
+            throw new InvalidOperationException(
+                $"The {channel} release advertises {manifest.Version} but carries no {manifest.Rpm}.");
+        }
+
+        found = (manifest.Version, manifest.RpmVersion, rpm, manifest.Sha256);
+
+        return new AvailableUpdate(manifest.Version);
+    }
+
+    /// <inheritdoc />
+    public async Task DownloadAsync(
+        AvailableUpdate update,
+        IProgress<int> progress,
+        CancellationToken cancellationToken)
+    {
+        ArgumentNullException.ThrowIfNull(update);
+        ArgumentNullException.ThrowIfNull(progress);
+
+        if (found is not { } target
+            || !string.Equals(target.Version, update.Version, StringComparison.Ordinal))
+        {
+            throw new InvalidOperationException(
+                "That update did not come from this channel's last check. Check again before downloading.");
+        }
+
+        // Under the profile directory rather than /tmp, because a package this size on a tmpfs is a
+        // machine that runs out of memory downloading an update. One name, overwritten, so a client that
+        // fetched three updates it never installed is not holding three hundred megabytes.
+        var directory = Path.Combine(dataDirectory, "updates");
+
+        Directory.CreateDirectory(directory);
+
+        var path = Path.Combine(directory, $"{packageName}-update.rpm");
+
+        await FetchAsync(target.Url, path, progress, cancellationToken).ConfigureAwait(false);
+
+        // ◆ Verified before it is ever named to the installer, and the file is removed if it does not
+        // match. Leaving a package that failed its digest on disk would put a file somebody could install
+        // by hand next to a filename that says it is an update.
+        var digest = Convert.ToHexStringLower(await Sha256OfAsync(path, cancellationToken).ConfigureAwait(false));
+
+        if (!string.Equals(digest, target.Sha256, StringComparison.OrdinalIgnoreCase))
+        {
+            File.Delete(path);
+
+            throw new InvalidOperationException(
+                $"The downloaded package is not the one {channel} published: expected {target.Sha256}, "
+                + $"got {digest}.");
+        }
+
+        fetched = path;
+    }
+
+    /// <summary>Streams one URL to one path, reporting progress as it goes.</summary>
+    private async Task FetchAsync(
+        string url,
+        string path,
+        IProgress<int> progress,
+        CancellationToken cancellationToken)
+    {
+        using var response = await http
+            .GetAsync(url, HttpCompletionOption.ResponseHeadersRead, cancellationToken)
+            .ConfigureAwait(false);
+
+        response.EnsureSuccessStatusCode();
+
+        var total = response.Content.Headers.ContentLength ?? 0;
+
+        var stream = await response.Content.ReadAsStreamAsync(cancellationToken).ConfigureAwait(false);
+        var destination = File.Create(path);
+
+        await using (stream.ConfigureAwait(false))
+        await using (destination.ConfigureAwait(false))
+        {
+            var buffer = new byte[81920];
+            long copied = 0;
+            int read;
+
+            while ((read = await stream.ReadAsync(buffer, cancellationToken).ConfigureAwait(false)) > 0)
+            {
+                await destination.WriteAsync(buffer.AsMemory(0, read), cancellationToken)
+                    .ConfigureAwait(false);
+
+                copied += read;
+
+                if (total > 0)
+                {
+                    progress.Report((int)(copied * 100 / total));
+                }
+            }
+        }
+    }
+
+    /// <inheritdoc />
+    /// <remarks>
+    /// <para>
+    /// <b>This installs nothing and returns immediately.</b> It asks PackageKit — the freedesktop service
+    /// every RPM distribution runs — to install a local file, which raises a polkit prompt drawn by the
+    /// desktop rather than by this application. The user may refuse, and refusing is not a failure.
+    /// </para>
+    /// <para>
+    /// <b>The process is not restarted and must not be.</b> Replacing files under a running process is
+    /// ordinary on Linux: the old inodes stay mapped and this build keeps running until somebody closes
+    /// it. That is the right behaviour for an application that deliberately keeps shells open across a
+    /// lock — the same argument ADR 0013 makes for waiting for a restart the user asks for — and it is why
+    /// the name inherited from the interface reads oddly here. There is no equivalent of Android's
+    /// "the platform ends the process", so nothing ends it.
+    /// </para>
+    /// </remarks>
+    public void ApplyAndRestart(AvailableUpdate update)
+    {
+        ArgumentNullException.ThrowIfNull(update);
+
+        if (fetched is not { } path || !File.Exists(path))
+        {
+            throw new InvalidOperationException("There is no downloaded update to install.");
+        }
+
+        handToInstaller(path);
+    }
+
+    /// <summary>Which release this build's channel reads.</summary>
+    /// <remarks>
+    /// <c>releases/latest</c> skips prereleases, which keeps the nightly — published as one — out of the
+    /// release channel's answer even though both live on the same forge.
+    /// </remarks>
+    private string ReleaseUrl() =>
+        string.Equals(channel, "nightly", StringComparison.Ordinal)
+            ? $"{RepositoryApi}/releases/tags/nightly-desktop"
+            : $"{RepositoryApi}/releases/latest";
+
+    private static string? Asset(LinuxForgeRelease? release, string name) =>
+        release?.Assets?.FirstOrDefault(
+            asset => string.Equals(asset.Name, name, StringComparison.Ordinal))?.DownloadUrl;
+
+    private async Task<T?> ReadAsync<T>(
+        string url,
+        JsonTypeInfo<T> shape,
+        CancellationToken cancellationToken)
+    {
+        var stream = await http.GetStreamAsync(url, cancellationToken).ConfigureAwait(false);
+
+        await using (stream.ConfigureAwait(false))
+        {
+            return await JsonSerializer
+                .DeserializeAsync(stream, shape, cancellationToken)
+                .ConfigureAwait(false);
+        }
+    }
+
+    private static async Task<byte[]> Sha256OfAsync(string path, CancellationToken cancellationToken)
+    {
+        var stream = File.OpenRead(path);
+
+        await using (stream.ConfigureAwait(false))
+        {
+            return await SHA256.HashDataAsync(stream, cancellationToken).ConfigureAwait(false);
+        }
+    }
+
+    /// <summary>
+    /// What rpm records for the installed package, which is the number a newer build must beat.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// Read from the package database rather than from the assembly, for the reason
+    /// <c>AndroidUpdateChannel.ReadInstalledVersionCode</c> gives about reading it from the platform: the
+    /// assembly's version is a SemVer string and this comparison has to be the one the installer will
+    /// make. Deriving the rpm version from the assembly here would mean a second copy of the translation
+    /// <c>scripts/ci-linux-rpm.sh</c> performs, and two copies of one rule is a pair that can disagree.
+    /// </para>
+    /// <para>
+    /// A machine that cannot answer gets null, which the caller reads as <c>0</c> and makes every
+    /// published build look newer. That is the wrong way to fail, chosen deliberately: the failure is then
+    /// an install PackageKit declines rather than a security fix that never arrives.
+    /// </para>
+    /// </remarks>
+    private static string? InstalledRpmVersion(string packageName)
+    {
+        try
+        {
+            using var process = Process.Start(new ProcessStartInfo("rpm")
+            {
+                ArgumentList = { "-q", "--queryformat", "%{VERSION}-%{RELEASE}", packageName },
+                RedirectStandardOutput = true,
+                RedirectStandardError = true,
+                UseShellExecute = false,
+            });
+
+            if (process is null)
+            {
+                return null;
+            }
+
+            var answer = process.StandardOutput.ReadToEnd().Trim();
+
+            process.WaitForExit(TimeSpan.FromSeconds(10));
+
+            return process.ExitCode == 0 && answer.Length > 0 ? answer : null;
+        }
+        catch (Exception exception) when (exception is not OutOfMemoryException)
+        {
+            return null;
+        }
+    }
+
+    /// <summary>Asks PackageKit to install a local file, and lets polkit do the asking.</summary>
+    /// <remarks>
+    /// <para>
+    /// <b>Over <c>gdbus</c> rather than a D-Bus library, and that is a dependency argument rather than a
+    /// stylistic one.</b> <c>gdbus</c> ships in glib2, which this package already depends on twice over —
+    /// GTK and WebKitGTK are the terminal — so it is present on every machine that can run this at all.
+    /// Adding a managed D-Bus stack to reach two method calls would be a package, a lock file entry and a
+    /// second thing to keep current, for no capability. <c>BrowserLauncher</c> settles the same trade the
+    /// same way one layer down.
+    /// </para>
+    /// <para>
+    /// <b>Two calls, because PackageKit's interface is a factory.</b> <c>CreateTransaction</c> returns an
+    /// object path and the work is asked of that; a transaction is single-use, so this cannot be cached.
+    /// </para>
+    /// <para>
+    /// <b>The flags word is zero, and zero is load-bearing.</b> Bit 0 is <c>ONLY_TRUSTED</c>. This
+    /// package is not signed — ADR 0013 rule 11 says why, and why that is not decision 8's "unsigned for
+    /// now" — so requiring trust would make every update fail at the last step. Clearing it is what routes
+    /// the request to polkit's <c>package-install-untrusted</c> action, which asks for an administrator
+    /// rather than proceeding quietly. The user sees what they are agreeing to.
+    /// </para>
+    /// </remarks>
+    private static void HandToPackageKit(string path)
+    {
+        var transaction = Gdbus(
+            "--system",
+            "--dest", "org.freedesktop.PackageKit",
+            "--object-path", "/org/freedesktop/PackageKit",
+            "--method", "org.freedesktop.PackageKit.CreateTransaction");
+
+        // gdbus prints a GVariant tuple — (objectpath '/17_beadcfae',) — and the path is the only thing
+        // in it. Parsed by taking what is between the quotes rather than with a regular expression over a
+        // format nobody has promised to keep.
+        var opened = transaction.IndexOf('\'', StringComparison.Ordinal);
+        var closed = transaction.LastIndexOf('\'');
+
+        if (opened < 0 || closed <= opened)
+        {
+            throw new InvalidOperationException(
+                $"PackageKit did not return a transaction to install into: {transaction}");
+        }
+
+        var objectPath = transaction[(opened + 1)..closed];
+
+        Gdbus(
+            "--system",
+            "--dest", "org.freedesktop.PackageKit",
+            "--object-path", objectPath,
+            "--method", "org.freedesktop.PackageKit.Transaction.InstallFiles",
+            "0",
+            $"['{path}']");
+    }
+
+    /// <summary>Runs one gdbus call and answers with its output.</summary>
+    private static string Gdbus(params string[] arguments)
+    {
+        var start = new ProcessStartInfo("gdbus")
+        {
+            RedirectStandardOutput = true,
+            RedirectStandardError = true,
+            UseShellExecute = false,
+        };
+
+        foreach (var argument in arguments)
+        {
+            start.ArgumentList.Add(argument);
+        }
+
+        using var process = Process.Start(start)
+            ?? throw new InvalidOperationException(
+                "gdbus could not be started, so the package could not be handed to PackageKit.");
+
+        var output = process.StandardOutput.ReadToEnd();
+        var error = process.StandardError.ReadToEnd();
+
+        process.WaitForExit(TimeSpan.FromSeconds(30));
+
+        if (process.ExitCode != 0)
+        {
+            // The message is shown to the user by UpdateViewModel, so it says what to do rather than what
+            // failed. A missing PackageKit is the likely cause on a minimal install and is the one thing
+            // they can fix.
+            throw new InvalidOperationException(
+                string.Format(
+                    CultureInfo.CurrentCulture,
+                    "PackageKit would not take the package. Install it with `sudo dnf install PackageKit`, "
+                    + "or install the file yourself: {0}. ({1})",
+                    arguments[^1],
+                    error.Trim()));
+        }
+
+        return output;
+    }
+}

src/DodoSSH.Client.App/Platform/MacDeviceKeyStore.cs

@@ -0,0 +1,598 @@
+using System.Runtime.InteropServices;
+using System.Runtime.Versioning;
+using System.Text;
+using DodoSSH.Client.Session;
+using static DodoSSH.Client.App.Platform.MacSecurity;
+
+namespace DodoSSH.Client.App.Platform;
+
+/// <summary>
+/// Keeps the device key encrypted to a Secure Enclave key whose use requires the user's presence.
+/// </summary>
+/// <remarks>
+/// <para>
+/// The macOS counterpart of <see cref="WindowsDeviceKeyStore"/>, and the same argument holds it up:
+/// <b>the consent is enforced by the platform, not by this class</b>. The unwrapping key is generated
+/// inside the Secure Enclave and never leaves it — there is no code path, privileged or otherwise, that
+/// turns it into bytes — and it is created under an access control requiring
+/// <see cref="AccessControlFlags.UserPresence"/>, so Touch ID or the login password is a condition of
+/// <em>using</em> it. Malware running as the user can ask for a decryption; it cannot answer the prompt,
+/// and the attempt is visible.
+/// </para>
+/// <para>
+/// A store that showed its own prompt and then read a protected file would be trivially bypassed, which
+/// is the mistake ADR 0007 originally described and the Windows store's comment corrects. The correction
+/// applies here unchanged.
+/// </para>
+/// <para>
+/// <b>P-256 and ECIES, where Windows uses RSA-OAEP, and the difference is not a preference.</b> The
+/// Secure Enclave holds exactly one kind of key: a 256-bit key on the NIST P-256 curve. It will not hold
+/// an RSA key at any size. So the wrap is <c>eciesEncryptionCofactorX963SHA256AESGCM</c> — an ephemeral
+/// agreement against the enclave's public half, X9.63-KDF to an AES-GCM key, and the ephemeral public
+/// key carried in the output. The framework does all of that; what matters here is that the input is 32
+/// bytes and there is no size limit worth worrying about.
+/// </para>
+/// <para>
+/// <b>Sealing is silent and unsealing prompts, which is better than the Windows shape rather than merely
+/// different.</b> On Windows, <c>CngKey.Create</c> with <c>ProtectKey</c> raises a dialog at creation as
+/// well, because the policy means "protect this key with a PIN" and Windows sets that up there and then.
+/// Here <see cref="SecKeyCopyPublicKey"/> works on an enclave key without any prompt, so registering a
+/// device shows nothing and only unlock asks. <see cref="SaveAsync"/> is therefore not user-facing on
+/// this platform — but it is still called from where the Windows one has to be, and relying on that
+/// difference would make the shared caller platform-specific for no gain.
+/// </para>
+/// <para>
+/// <b>What this cannot be tested against, and what follows from that.</b> Every method except
+/// <see cref="IsAvailableAsync"/> and the empty case of <see cref="TryLoadAsync"/> needs an interactive
+/// login session and real enclave hardware, so none can be exercised by an automated test — the same
+/// line the Windows store draws. It also means <see cref="IsSupported"/> must probe rather than infer:
+/// see its remarks for the three ordinary machines that have no usable enclave and must degrade to the
+/// passphrase rather than fail at unlock.
+/// </para>
+/// </remarks>
+[SupportedOSPlatform("macos")]
+public sealed partial class MacDeviceKeyStore : IDeviceKeyStore
+{
+    /// <summary>
+    /// The keychain tag this application's enclave key is filed under.
+    /// </summary>
+    /// <remarks>
+    /// Versioned for the reason the Windows key name is: a future change of curve or wrap algorithm can
+    /// create a new key beside the old one rather than failing to open blobs written by a previous
+    /// build. A device that cannot be opened falls back to the passphrase, which is survivable — but
+    /// silently, and a user would only notice their fingerprint had stopped working.
+    ///
+    /// Prefixed with the bundle identifier because the keychain is shared across every application the
+    /// user runs, unlike a CNG key name, which is scoped to the user's key store already.
+    /// </remarks>
+    private const string KeyTag = "dev.dodotech.dodossh.devicekey.v1";
+
+    /// <summary>
+    /// Shown in the Touch ID prompt, so it has to read as a sentence to a person.
+    /// </summary>
+    /// <remarks>
+    /// macOS composes it into "DodoSSH is trying to ...", so this is a verb phrase and not a sentence of
+    /// its own. The same words the Windows consent dialog uses.
+    /// </remarks>
+    private const string ConsentPrompt = "unlock your DodoSSH vault";
+
+    private readonly ClientPaths paths;
+
+    /// <summary>Creates the store.</summary>
+    public MacDeviceKeyStore(ClientPaths paths)
+    {
+        ArgumentNullException.ThrowIfNull(paths);
+        this.paths = paths;
+    }
+
+    /// <summary>
+    /// Whether this Mac has a Secure Enclave that will hold a key for this build.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// Probed by creating a throwaway key and deleting it, rather than by asking whether the hardware
+    /// exists. Three ordinary situations answer "no" here and would otherwise only be discovered at the
+    /// moment somebody tried to unlock:
+    /// </para>
+    /// <para>
+    /// <b>An Intel Mac with no T2.</b> Apple Silicon and T2 machines have an enclave; earlier Intel
+    /// models do not, and there is no single attribute that says so.
+    /// </para>
+    /// <para>
+    /// <b>A build that is not code signed.</b> Enclave key creation requires a signing identity, so
+    /// every <c>dotnet run</c> and every build from an IDE fails here with a missing-entitlement error.
+    /// That is the correct answer rather than a nuisance: a development build should keep asking for the
+    /// passphrase, and this is what makes it do so without a platform check somewhere else.
+    /// </para>
+    /// <para>
+    /// <b>A machine with no login password set.</b> <see cref="AccessControlFlags.UserPresence"/> has
+    /// nothing to demand, and the framework refuses the access control object rather than silently
+    /// creating a key anybody could use.
+    /// </para>
+    /// <para>
+    /// The probe uses its own tag and no UI policy, so nothing prompts and nothing collides with the
+    /// real key. It is deleted immediately; a probe key left behind would accumulate one per launch.
+    /// </para>
+    /// </remarks>
+    internal static bool IsSupported()
+    {
+        try
+        {
+            var probe = $"{KeyTag}.probe.{Guid.CreateVersion7():N}";
+
+            using var scope = new CoreFoundationScope();
+
+            var symbols = MacSymbols.Resolve();
+
+            if (!symbols.Complete)
+            {
+                return false;
+            }
+
+            var key = CreateEnclaveKey(scope, symbols, probe);
+
+            if (key == IntPtr.Zero)
+            {
+                return false;
+            }
+
+            // Discarded deliberately. The question this method answers is whether the enclave will make a
+            // key, and it demonstrably just did; a failure to clean the probe up afterwards leaves one
+            // stray keychain item and does not make the answer no.
+            _ = DeleteKey(symbols, probe);
+
+            return true;
+        }
+        catch (Exception exception) when (exception is DllNotFoundException
+                                              or EntryPointNotFoundException
+                                              or BadImageFormatException)
+        {
+            // A macOS without these frameworks is not a thing that exists, so this is really the guard
+            // for the case that does: a future release renaming or removing one of them. The answer is
+            // the same as for hardware that is absent — no device key, ask for the passphrase.
+            return false;
+        }
+    }
+
+    /// <inheritdoc />
+    public ValueTask<bool> IsAvailableAsync(CancellationToken cancellationToken) =>
+        ValueTask.FromResult(IsSupported());
+
+    /// <inheritdoc />
+    public async ValueTask SaveAsync(
+        ReadOnlyMemory<byte> devicePrivateKey,
+        CancellationToken cancellationToken)
+    {
+        var sealedKey = Seal(devicePrivateKey.Span)
+            ?? throw new InvalidOperationException(
+                "The Secure Enclave would not seal the device key. Check IsAvailableAsync before offering to register one.");
+
+        paths.EnsureCreated();
+
+        await File.WriteAllBytesAsync(paths.DeviceKeyFile, sealedKey, cancellationToken)
+            .ConfigureAwait(false);
+    }
+
+    /// <inheritdoc />
+    public async ValueTask<byte[]?> TryLoadAsync(CancellationToken cancellationToken)
+    {
+        if (!File.Exists(paths.DeviceKeyFile))
+        {
+            return null;
+        }
+
+        var sealedKey = await File.ReadAllBytesAsync(paths.DeviceKeyFile, cancellationToken)
+            .ConfigureAwait(false);
+
+        return Unseal(sealedKey);
+    }
+
+    /// <inheritdoc />
+    public ValueTask ForgetAsync(CancellationToken cancellationToken)
+    {
+        if (File.Exists(paths.DeviceKeyFile))
+        {
+            File.Delete(paths.DeviceKeyFile);
+        }
+
+        var symbols = MacSymbols.Resolve();
+
+        if (symbols.Complete)
+        {
+            // Discarded, and that is deliberate: there is nothing a caller could do about a failure here,
+            // and the file deleted above is the half that decides whether unlock will try at all. A key
+            // left in the enclave with no ciphertext to open is inert.
+            _ = DeleteKey(symbols, KeyTag);
+        }
+
+        return ValueTask.CompletedTask;
+    }
+
+    /// <remarks>
+    /// Silent: it uses only the public half. Null on every failure, and the caller's answer to all of
+    /// them is the same — do not offer a device unlock.
+    /// </remarks>
+    private static byte[]? Seal(ReadOnlySpan<byte> devicePrivateKey)
+    {
+        try
+        {
+            using var scope = new CoreFoundationScope();
+
+            var symbols = MacSymbols.Resolve();
+
+            if (!symbols.Complete)
+            {
+                return null;
+            }
+
+            // Created on first use rather than at registration, so that a device key re-registered after
+            // a ForgetAsync gets a key again without anything having to notice that it had gone.
+            var privateKey = FindKey(scope, symbols, KeyTag, prompt: null);
+
+            if (privateKey == IntPtr.Zero)
+            {
+                privateKey = CreateEnclaveKey(scope, symbols, KeyTag);
+            }
+
+            if (privateKey == IntPtr.Zero)
+            {
+                return null;
+            }
+
+            var publicKey = scope.Keep(SecKeyCopyPublicKey(privateKey));
+
+            if (publicKey == IntPtr.Zero)
+            {
+                return null;
+            }
+
+            var plaintext = Data(scope, devicePrivateKey);
+
+            if (plaintext == IntPtr.Zero)
+            {
+                return null;
+            }
+
+            var ciphertext = scope.Keep(
+                SecKeyCreateEncryptedData(publicKey, symbols.EciesAlgorithm, plaintext, out var error));
+
+            scope.Keep(error);
+
+            return ciphertext == IntPtr.Zero ? null : ToArray(ciphertext);
+        }
+        catch (Exception exception) when (exception is DllNotFoundException
+                                              or EntryPointNotFoundException
+                                              or BadImageFormatException)
+        {
+            return null;
+        }
+    }
+
+    /// <remarks>
+    /// <para>
+    /// This is the call that prompts. Every failure becomes null, and the set is wider than it looks:
+    /// the key may be gone, the user may have cancelled or let the prompt time out, the enclave may have
+    /// invalidated it after the login password was reset, or the blob may predate a key that has since
+    /// been replaced. None of them are distinguishable to a user and all have the same remedy, so none
+    /// are worth telling apart here — see <c>UnlockStatus.DeviceKeyUnavailable</c>.
+    /// </para>
+    /// <para>
+    /// Blocking, and it blocks on a person. The prompt is modal to the application, so this must not run
+    /// on a thread that is also expected to draw the window behind it.
+    /// </para>
+    /// </remarks>
+    private static byte[]? Unseal(byte[] sealedKey)
+    {
+        try
+        {
+            using var scope = new CoreFoundationScope();
+
+            var symbols = MacSymbols.Resolve();
+
+            if (!symbols.Complete)
+            {
+                return null;
+            }
+
+            var privateKey = FindKey(scope, symbols, KeyTag, ConsentPrompt);
+
+            if (privateKey == IntPtr.Zero)
+            {
+                return null;
+            }
+
+            var ciphertext = Data(scope, sealedKey);
+
+            if (ciphertext == IntPtr.Zero)
+            {
+                return null;
+            }
+
+            var plaintext = scope.Keep(
+                SecKeyCreateDecryptedData(privateKey, symbols.EciesAlgorithm, ciphertext, out var error));
+
+            scope.Keep(error);
+
+            return plaintext == IntPtr.Zero ? null : ToArray(plaintext);
+        }
+        catch (Exception exception) when (exception is DllNotFoundException
+                                              or EntryPointNotFoundException
+                                              or BadImageFormatException)
+        {
+            return null;
+        }
+    }
+
+    /// <summary>
+    /// Generates a key inside the Secure Enclave, filed under <paramref name="tag"/>. Owned by the scope.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// The attribute dictionary is the whole security decision, so it is worth reading rather than
+    /// pattern-matching. <c>TokenID = SecureEnclave</c> is what puts the private half in hardware;
+    /// without it this silently generates an ordinary software key that behaves identically in every
+    /// visible way and protects nothing.
+    /// </para>
+    /// <para>
+    /// <c>AccessibleWhenUnlockedThisDeviceOnly</c> rather than any of the migratable classes, because a
+    /// device key that could be restored onto another machine from a backup would no longer mean "this
+    /// machine". The enclave already makes that impossible; saying it as well means the intent survives
+    /// a future change of storage.
+    /// </para>
+    /// <para>
+    /// <c>UseDataProtectionKeychain</c> is the macOS-specific one and the easiest to omit. Without it,
+    /// macOS routes this to the older file-based keychain, which does not understand access control
+    /// objects or the enclave, and the call fails with a parameter error that says nothing about the
+    /// missing key.
+    /// </para>
+    /// </remarks>
+    private static IntPtr CreateEnclaveKey(CoreFoundationScope scope, MacSymbols symbols, string tag)
+    {
+        var access = scope.Keep(SecAccessControlCreateWithFlags(
+            IntPtr.Zero,
+            symbols.AccessibleWhenUnlockedThisDeviceOnly,
+            AccessControlFlags.PrivateKeyUsage | AccessControlFlags.UserPresence,
+            out var accessError));
+
+        scope.Keep(accessError);
+
+        if (access == IntPtr.Zero)
+        {
+            return IntPtr.Zero;
+        }
+
+        var privateAttrs = Dictionary(
+            scope,
+            [symbols.AttrIsPermanent, symbols.AttrApplicationTag, symbols.AttrAccessControl],
+            [symbols.True, TagData(scope, tag), access]);
+
+        if (privateAttrs == IntPtr.Zero)
+        {
+            return IntPtr.Zero;
+        }
+
+        var keySize = Number(scope, 256);
+
+        var parameters = Dictionary(
+            scope,
+            [
+                symbols.AttrKeyType,
+                symbols.AttrKeySizeInBits,
+                symbols.AttrTokenId,
+                symbols.UseDataProtectionKeychain,
+                symbols.PrivateKeyAttrs,
+            ],
+            [
+                symbols.KeyTypeEcSecPrimeRandom,
+                keySize,
+                symbols.TokenIdSecureEnclave,
+                symbols.True,
+                privateAttrs,
+            ]);
+
+        if (parameters == IntPtr.Zero)
+        {
+            return IntPtr.Zero;
+        }
+
+        var key = scope.Keep(SecKeyCreateRandomKey(parameters, out var error));
+
+        scope.Keep(error);
+
+        return key;
+    }
+
+    /// <summary>
+    /// Looks the enclave key up by tag. Owned by the scope; zero when there is none.
+    /// </summary>
+    /// <remarks>
+    /// <paramref name="prompt"/> is attached here and consumed later: the lookup itself does not raise
+    /// anything, because a handle to an enclave key is not a use of it. The words reach the user at the
+    /// decrypt, which is the operation the access control actually guards.
+    ///
+    /// <c>UseOperationPrompt</c> is deprecated in favour of an <c>LAContext</c>, and is used anyway. An
+    /// LAContext would mean binding LocalAuthentication as well for one string, and the deprecated key
+    /// still works; the day it stops, this call fails and the store degrades to the passphrase, which is
+    /// the failure this whole class is built to degrade into.
+    /// </remarks>
+    private static IntPtr FindKey(CoreFoundationScope scope, MacSymbols symbols, string tag, string? prompt)
+    {
+        List<IntPtr> keys =
+        [
+            symbols.Class,
+            symbols.AttrApplicationTag,
+            symbols.AttrKeyType,
+            symbols.UseDataProtectionKeychain,
+            symbols.ReturnRef,
+        ];
+
+        List<IntPtr> values =
+        [
+            symbols.ClassKey,
+            TagData(scope, tag),
+            symbols.KeyTypeEcSecPrimeRandom,
+            symbols.True,
+            symbols.True,
+        ];
+
+        if (prompt is not null)
+        {
+            keys.Add(symbols.UseOperationPrompt);
+            values.Add(scope.Keep(CFString(prompt)));
+        }
+
+        var query = Dictionary(scope, [.. keys], [.. values]);
+
+        if (query == IntPtr.Zero)
+        {
+            return IntPtr.Zero;
+        }
+
+        var status = SecItemCopyMatching(query, out var result);
+
+        // errSecItemNotFound is the ordinary answer on a machine that has never registered a device, and
+        // it is not distinguished from any other failure for the reason the class remarks give.
+        return status == Success ? scope.Keep(result) : IntPtr.Zero;
+    }
+
+    /// <summary>Removes the key with this tag from the keychain.</summary>
+    /// <returns>Whether the keychain now has no key under this tag.</returns>
+    /// <remarks>
+    /// <c>ItemNotFound</c> counts as success, and that is the common case rather than an edge: it is
+    /// what a machine that never registered a device answers, and what the second of two
+    /// <see cref="ForgetAsync"/> calls answers. Treating it as a failure would make forgetting a device
+    /// twice report a problem that does not exist.
+    /// </remarks>
+    private static bool DeleteKey(MacSymbols symbols, string tag)
+    {
+        using var scope = new CoreFoundationScope();
+
+        var query = Dictionary(
+            scope,
+            [symbols.Class, symbols.AttrApplicationTag, symbols.UseDataProtectionKeychain],
+            [symbols.ClassKey, TagData(scope, tag), symbols.True]);
+
+        if (query == IntPtr.Zero)
+        {
+            return false;
+        }
+
+        var status = SecItemDelete(query);
+
+        return status is Success or ItemNotFound;
+    }
+
+    // ---- Small CoreFoundation conveniences ---------------------------------------------------------
+
+    /// <remarks>
+    /// The arrays are pinned for the duration of the call and not beyond it, which is correct because
+    /// <c>CFDictionaryCreate</c> copies them: the dictionary retains each key and value, and never reads
+    /// the arrays again.
+    /// </remarks>
+    private static IntPtr Dictionary(CoreFoundationScope scope, IntPtr[] keys, IntPtr[] values)
+    {
+        // A zero anywhere means one of the constants did not resolve or an earlier allocation failed.
+        // Passing it on produces a dictionary with a null key, which CFDictionaryCreate does not reject
+        // — it crashes inside the callback table instead.
+        if (Array.IndexOf(keys, IntPtr.Zero) >= 0 || Array.IndexOf(values, IntPtr.Zero) >= 0)
+        {
+            return IntPtr.Zero;
+        }
+
+        var symbols = MacSymbols.Resolve();
+
+        unsafe
+        {
+            fixed (IntPtr* keyPtr = keys)
+            fixed (IntPtr* valuePtr = values)
+            {
+                return scope.Keep(CFDictionaryCreate(
+                    IntPtr.Zero,
+                    (IntPtr)keyPtr,
+                    (IntPtr)valuePtr,
+                    keys.Length,
+                    symbols.TypeDictionaryKeyCallBacks,
+                    symbols.TypeDictionaryValueCallBacks));
+            }
+        }
+    }
+
+    /// <summary>Copies bytes into a CFData. Owned by the scope.</summary>
+    /// <remarks>
+    /// The pin lasts only as long as the call, which is correct: <c>CFDataCreate</c> copies, so the
+    /// CFData does not reference this memory afterwards. <c>CFDataCreateWithBytesNoCopy</c> would not,
+    /// and is not used for exactly that reason — it would hand the framework a pointer into the managed
+    /// heap and rely on the object staying where the collector first put it.
+    /// </remarks>
+    private static IntPtr Data(CoreFoundationScope scope, ReadOnlySpan<byte> bytes)
+    {
+        unsafe
+        {
+            fixed (byte* pointer = bytes)
+            {
+                return scope.Keep(CFDataCreate(IntPtr.Zero, (IntPtr)pointer, bytes.Length));
+            }
+        }
+    }
+
+    /// <remarks>
+    /// UTF-8 rather than any other encoding, and it only has to be consistent with itself: the tag is an
+    /// opaque blob the keychain matches byte for byte, so what matters is that a lookup encodes it the
+    /// same way the creation did. It is written once, here, for exactly that reason.
+    /// </remarks>
+    private static IntPtr TagData(CoreFoundationScope scope, string tag) =>
+        Data(scope, Encoding.UTF8.GetBytes(tag));
+
+    private static IntPtr Number(CoreFoundationScope scope, int value)
+    {
+        unsafe
+        {
+            return scope.Keep(CFNumberCreate(IntPtr.Zero, (nint)CFNumberIntType, (IntPtr)(&value)));
+        }
+    }
+
+    /// <summary>Builds a CFString from a managed string. Owned, so the caller tracks it.</summary>
+    /// <remarks>
+    /// Built explicitly rather than left to the marshaller, because these calls take a
+    /// <c>CFStringRef</c> and not a C string — the runtime's default marshalling would hand over a
+    /// <c>char*</c>, which CoreFoundation reads as an object pointer and follows into nothing.
+    /// </remarks>
+    private static IntPtr CFString(string value)
+    {
+        var bytes = Encoding.UTF8.GetBytes(value);
+
+        unsafe
+        {
+            fixed (byte* pointer = bytes)
+            {
+                // kCFStringEncodingUTF8 is 0x08000100, spelled out rather than named because it is the
+                // only encoding constant this file uses.
+                return CFStringCreateWithBytes(IntPtr.Zero, (IntPtr)pointer, bytes.Length, 0x08000100, false);
+            }
+        }
+    }
+
+    [LibraryImport(CoreFoundation)]
+    private static partial IntPtr CFStringCreateWithBytes(
+        IntPtr allocator,
+        IntPtr bytes,
+        nint numBytes,
+        uint encoding,
+        [MarshalAs(UnmanagedType.U1)] bool isExternalRepresentation);
+
+    private static byte[] ToArray(IntPtr data)
+    {
+        var length = (int)CFDataGetLength(data);
+        var pointer = CFDataGetBytePtr(data);
+
+        if (length <= 0 || pointer == IntPtr.Zero)
+        {
+            return [];
+        }
+
+        var result = new byte[length];
+        Marshal.Copy(pointer, result, 0, length);
+
+        return result;
+    }
+}

src/DodoSSH.Client.App/Platform/MacSecurity.cs

@@ -0,0 +1,254 @@
+using System.Runtime.InteropServices;
+using System.Runtime.Versioning;
+
+namespace DodoSSH.Client.App.Platform;
+
+/// <summary>
+/// The pieces of CoreFoundation and Security.framework <see cref="MacDeviceKeyStore"/> needs.
+/// </summary>
+/// <remarks>
+/// <para>
+/// Separated from the store itself because it is a different kind of code with a different kind of
+/// review: nothing here makes a decision, and everything here is a translation of a C declaration that
+/// is either right or wrong. Mixing the two would mean the security argument in
+/// <see cref="MacDeviceKeyStore"/> had to be read past two hundred lines of marshalling to find.
+/// </para>
+/// <para>
+/// <b>Every Create or Copy returns an object this process owns.</b> That is CoreFoundation's Create
+/// Rule, and it is the thing here that goes wrong silently: the enclave key handle is small, so a leak
+/// shows up as nothing at all until a long-running process has done a few thousand unlocks.
+/// <see cref="CoreFoundationScope"/> exists so ownership is tracked by construction rather than by
+/// remembering, and every function below that returns a handle says whether it is owned.
+/// </para>
+/// <para>
+/// <b>The integer widths are the part worth checking against the headers rather than skimming.</b>
+/// <c>CFIndex</c>, <c>CFOptionFlags</c> and <c>CFNumberType</c> are all pointer-width on a 64-bit Mac,
+/// not 32-bit, and getting one wrong does not fail cleanly — it shifts every argument after it, so the
+/// call receives plausible rubbish and returns a parameter error that names nothing.
+/// </para>
+/// </remarks>
+[SupportedOSPlatform("macos")]
+internal static partial class MacSecurity
+{
+    internal const string SecurityFramework =
+        "/System/Library/Frameworks/Security.framework/Security";
+
+    internal const string CoreFoundation =
+        "/System/Library/Frameworks/CoreFoundation.framework/CoreFoundation";
+
+    /// <summary>
+    /// The access control flags <c>SecAccessControlCreateWithFlags</c> takes.
+    /// </summary>
+    /// <remarks>
+    /// <c>ulong</c> because the parameter is a <c>CFOptionFlags</c>, which is an <c>unsigned long</c>.
+    /// Only the two flags that are used are listed; the full set is large, and copying it in would
+    /// invite somebody to reach for one without reading what it does to the prompt — <c>Biometry</c>
+    /// alone, for instance, leaves a Mac with no Touch ID unable to unlock at all rather than falling
+    /// back to the login password.
+    /// </remarks>
+    [Flags]
+    internal enum AccessControlFlags : ulong
+    {
+        /// <summary>
+        /// Touch ID if the machine has it, the login password if not.
+        /// </summary>
+        /// <remarks>
+        /// The forgiving one, deliberately. <c>BiometryCurrentSet</c> would additionally invalidate the
+        /// key whenever a fingerprint is added or removed, which sounds stricter and here buys nothing:
+        /// this key wraps a device key whose loss already means "ask for the passphrase", so the only
+        /// effect would be users being sent back to their passphrase by an unrelated Settings change
+        /// they would never connect to it.
+        /// </remarks>
+        UserPresence = 1ul << 0,
+
+        /// <summary>Required for any key that lives in the Secure Enclave.</summary>
+        PrivateKeyUsage = 1ul << 30,
+    }
+
+    /// <summary>The CFNumberType code for a 32-bit int, from CFNumber.h.</summary>
+    internal const long CFNumberIntType = 9;
+
+    /// <summary>errSecSuccess.</summary>
+    internal const int Success = 0;
+
+    /// <summary>errSecItemNotFound, which is an answer rather than a failure.</summary>
+    internal const int ItemNotFound = -25300;
+
+    // ---- CoreFoundation ------------------------------------------------------------------------------
+
+    /// <summary>Releases an owned handle.</summary>
+    [LibraryImport(CoreFoundation)]
+    internal static partial void CFRelease(IntPtr handle);
+
+    /// <summary>Copies bytes into a new CFData. Owned.</summary>
+    [LibraryImport(CoreFoundation)]
+    internal static partial IntPtr CFDataCreate(IntPtr allocator, IntPtr bytes, nint length);
+
+    [LibraryImport(CoreFoundation)]
+    internal static partial IntPtr CFDataGetBytePtr(IntPtr data);
+
+    [LibraryImport(CoreFoundation)]
+    internal static partial nint CFDataGetLength(IntPtr data);
+
+    /// <summary>Boxes a value as a CFNumber. Owned.</summary>
+    [LibraryImport(CoreFoundation)]
+    internal static partial IntPtr CFNumberCreate(IntPtr allocator, nint theType, IntPtr valuePtr);
+
+    /// <summary>Builds an immutable dictionary. Owned.</summary>
+    /// <remarks>
+    /// <para>
+    /// The key and value arrays are passed as raw pointers to memory the caller pins, rather than as
+    /// managed arrays. Source-generated interop wants an explicit element count for a marshalled array,
+    /// and supplying one here would mean stating the length twice — once for the marshaller and once as
+    /// <paramref name="numValues"/> — which is exactly the pair that drifts.
+    /// </para>
+    /// <para>
+    /// The two callback tables are what make the dictionary retain its keys and values, which is why
+    /// they are passed rather than left null: with null callbacks the dictionary stores raw pointers and
+    /// keeps nothing alive, and the resulting use-after-free is intermittent by nature.
+    /// </para>
+    /// </remarks>
+    [LibraryImport(CoreFoundation)]
+    internal static partial IntPtr CFDictionaryCreate(
+        IntPtr allocator,
+        IntPtr keys,
+        IntPtr values,
+        nint numValues,
+        IntPtr keyCallBacks,
+        IntPtr valueCallBacks);
+
+    // ---- Security.framework --------------------------------------------------------------------------
+
+    /// <summary>Builds the access policy a Secure Enclave key is created under. Owned.</summary>
+    [LibraryImport(SecurityFramework)]
+    internal static partial IntPtr SecAccessControlCreateWithFlags(
+        IntPtr allocator,
+        IntPtr protection,
+        AccessControlFlags flags,
+        out IntPtr error);
+
+    /// <summary>Creates a key pair from an attribute dictionary. Owned.</summary>
+    [LibraryImport(SecurityFramework)]
+    internal static partial IntPtr SecKeyCreateRandomKey(IntPtr parameters, out IntPtr error);
+
+    /// <summary>The public half of a key. Owned.</summary>
+    /// <remarks>
+    /// Available even for an enclave key, and that asymmetry is the whole reason this design works: the
+    /// public half is an ordinary key this process can hold and use, while the private half is a handle
+    /// to something inside the enclave that never becomes bytes. So sealing is silent and unsealing is
+    /// the thing the user is asked about.
+    /// </remarks>
+    [LibraryImport(SecurityFramework)]
+    internal static partial IntPtr SecKeyCopyPublicKey(IntPtr key);
+
+    /// <summary>Encrypts with a public key. Owned.</summary>
+    [LibraryImport(SecurityFramework)]
+    internal static partial IntPtr SecKeyCreateEncryptedData(
+        IntPtr key,
+        IntPtr algorithm,
+        IntPtr plaintext,
+        out IntPtr error);
+
+    /// <summary>Decrypts with a private key, prompting for whatever guards it. Owned.</summary>
+    [LibraryImport(SecurityFramework)]
+    internal static partial IntPtr SecKeyCreateDecryptedData(
+        IntPtr key,
+        IntPtr algorithm,
+        IntPtr ciphertext,
+        out IntPtr error);
+
+    /// <summary>Finds a keychain item. The out handle is owned when the result is <see cref="Success"/>.</summary>
+    [LibraryImport(SecurityFramework)]
+    internal static partial int SecItemCopyMatching(IntPtr query, out IntPtr result);
+
+    /// <summary>Deletes every keychain item matching the query.</summary>
+    [LibraryImport(SecurityFramework)]
+    internal static partial int SecItemDelete(IntPtr query);
+
+    // ---- The framework constants ---------------------------------------------------------------------
+
+    /// <summary>
+    /// Reads one of a framework's global CFString constants, or zero if it is not exported.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// The keys these dictionaries take are not strings this code may spell for itself. They are
+    /// pointer-comparable constants exported by the framework, and a CFString built here with the same
+    /// characters is a different object — the lookups would miss and the call would fail with a
+    /// parameter error naming nothing.
+    /// </para>
+    /// <para>
+    /// <b>Dereferenced once, because the exported symbol is the variable rather than its value.</b>
+    /// <c>TryGetExport</c> answers the address of the global; the CFStringRef is what that address
+    /// holds. Missing the indirection produces a pointer that is stable, plausible and wrong, which is
+    /// the worst of the three available outcomes.
+    /// </para>
+    /// <para>
+    /// Zero on a missing symbol rather than an exception, because the caller's answer to every failure
+    /// is the same one — report the store unavailable and let unlock ask for the passphrase — and a
+    /// constant that has been renamed by a future macOS should reach that answer rather than a crash.
+    /// </para>
+    /// </remarks>
+    internal static IntPtr Constant(IntPtr library, string symbol) =>
+        NativeLibrary.TryGetExport(library, symbol, out var address)
+            ? Marshal.ReadIntPtr(address)
+            : IntPtr.Zero;
+}
+
+/// <summary>
+/// Releases every CoreFoundation handle put into it, in reverse order, exactly once.
+/// </summary>
+/// <remarks>
+/// <para>
+/// The alternative is a try/finally per handle, and the operations here need six or seven at a time — a
+/// dictionary holding a nested dictionary holding an access control object holding a CFData tag. Finallys
+/// nested that deep stop being read, and a handle released twice is a crash rather than a leak.
+/// </para>
+/// <para>
+/// <see cref="Keep"/> returns what it was given, so a handle can be tracked in the same expression that
+/// produces it and the call sites read as ordinary code.
+/// </para>
+/// </remarks>
+[SupportedOSPlatform("macos")]
+internal sealed class CoreFoundationScope : IDisposable
+{
+    private readonly List<IntPtr> owned = [];
+
+    private bool disposed;
+
+    /// <summary>Takes ownership of a handle and hands it straight back.</summary>
+    /// <remarks>
+    /// Zero is ignored rather than rejected. Every CoreFoundation call here answers zero on failure, so
+    /// accepting it lets a caller track the result in the expression that produces it and check it on
+    /// the next line, instead of writing the check twice.
+    /// </remarks>
+    internal IntPtr Keep(IntPtr handle)
+    {
+        if (handle != IntPtr.Zero)
+        {
+            owned.Add(handle);
+        }
+
+        return handle;
+    }
+
+    public void Dispose()
+    {
+        if (disposed)
+        {
+            return;
+        }
+
+        disposed = true;
+
+        // Reverse order, so a container is released before the things it retains. CoreFoundation does not
+        // require it — retain counts make the order irrelevant — but it keeps the lifetimes readable in a
+        // debugger, where a released container that still lists its contents is a confusing thing to meet.
+        for (var i = owned.Count - 1; i >= 0; i--)
+        {
+            MacSecurity.CFRelease(owned[i]);
+        }
+
+        owned.Clear();
+    }
+}

src/DodoSSH.Client.App/Platform/MacSymbols.cs

@@ -0,0 +1,185 @@
+using System.Runtime.InteropServices;
+using System.Runtime.Versioning;
+
+namespace DodoSSH.Client.App.Platform;
+
+/// <summary>
+/// The framework constants <see cref="MacDeviceKeyStore"/> passes to CoreFoundation and Security.
+/// </summary>
+/// <remarks>
+/// <para>
+/// Every field here is a pointer read out of a loaded framework rather than a value this code could
+/// write down. The dictionaries these go into are matched by pointer identity, so a CFString built with
+/// the same characters is a different key and the lookup misses — see <see cref="MacSecurity.Constant"/>
+/// for the indirection that trips people.
+/// </para>
+/// <para>
+/// <b>Resolved once and cached, and the caching is what makes the failure survivable.</b> Two frameworks
+/// and nineteen symbols is a lot of things to be wrong about, and the useful property is that being
+/// wrong about any one of them shows up here — as <see cref="Complete"/> being false — rather than
+/// three calls later as a parameter error. A store that reports itself unavailable sends the user back
+/// to their passphrase; a store that half works corrupts the moment somebody registers a device.
+/// </para>
+/// <para>
+/// <see cref="Lazy{T}"/> rather than a static constructor, because a type initialiser that throws
+/// poisons the type for the life of the process and turns a missing symbol into a
+/// <c>TypeInitializationException</c> at every later call site. The load is done inside a try instead,
+/// and its failure is a value.
+/// </para>
+/// </remarks>
+[SupportedOSPlatform("macos")]
+internal sealed class MacSymbols
+{
+    private static readonly Lazy<MacSymbols> Cached = new(Load, LazyThreadSafetyMode.ExecutionAndPublication);
+
+    private MacSymbols()
+    {
+    }
+
+    /// <summary>Whether every symbol resolved.</summary>
+    /// <remarks>
+    /// Checked by every caller before any of the pointers are used. It is one check rather than
+    /// nineteen, which is the only reason the call sites in <see cref="MacDeviceKeyStore"/> are
+    /// readable.
+    ///
+    /// Computed rather than stored, so that the instance returned when a framework will not load at all
+    /// — every field left at zero — answers false without that having to be set anywhere. One rule,
+    /// applied to the only state there is.
+    /// </remarks>
+    internal bool Complete => AllResolved();
+
+    // CoreFoundation.
+    internal IntPtr True { get; private init; }
+
+    internal IntPtr TypeDictionaryKeyCallBacks { get; private init; }
+
+    internal IntPtr TypeDictionaryValueCallBacks { get; private init; }
+
+    // Security: item classes and query keys.
+    internal IntPtr Class { get; private init; }
+
+    internal IntPtr ClassKey { get; private init; }
+
+    internal IntPtr ReturnRef { get; private init; }
+
+    internal IntPtr UseDataProtectionKeychain { get; private init; }
+
+    internal IntPtr UseOperationPrompt { get; private init; }
+
+    // Security: key attributes.
+    internal IntPtr AttrKeyType { get; private init; }
+
+    internal IntPtr AttrKeySizeInBits { get; private init; }
+
+    internal IntPtr AttrTokenId { get; private init; }
+
+    internal IntPtr AttrIsPermanent { get; private init; }
+
+    internal IntPtr AttrApplicationTag { get; private init; }
+
+    internal IntPtr AttrAccessControl { get; private init; }
+
+    internal IntPtr PrivateKeyAttrs { get; private init; }
+
+    // Security: attribute values.
+    internal IntPtr KeyTypeEcSecPrimeRandom { get; private init; }
+
+    internal IntPtr TokenIdSecureEnclave { get; private init; }
+
+    internal IntPtr AccessibleWhenUnlockedThisDeviceOnly { get; private init; }
+
+    /// <summary>
+    /// <c>kSecKeyAlgorithmECIESEncryptionCofactorX963SHA256AESGCM</c>.
+    /// </summary>
+    /// <remarks>
+    /// The one algorithm the Secure Enclave's P-256 keys support for encryption, and the reason this
+    /// store wraps rather than signs. The long name spells out the whole construction: an ephemeral
+    /// key agreed against the enclave's public half with cofactor ECDH, run through the X9.63 KDF with
+    /// SHA-256, used as an AES-GCM key. The ephemeral public key travels in the output, which is why the
+    /// ciphertext is larger than the 32 bytes going in and why nothing else has to be stored beside it.
+    /// </remarks>
+    internal IntPtr EciesAlgorithm { get; private init; }
+
+    /// <summary>The resolved symbols, loaded once.</summary>
+    internal static MacSymbols Resolve() => Cached.Value;
+
+    private static MacSymbols Load()
+    {
+        try
+        {
+            if (!NativeLibrary.TryLoad(MacSecurity.CoreFoundation, out var cf)
+                || !NativeLibrary.TryLoad(MacSecurity.SecurityFramework, out var sec))
+            {
+                // Every pointer left at zero, which AllResolved reads as incomplete.
+                return new MacSymbols();
+            }
+
+            // The two callback tables are structs rather than object pointers, so what is wanted is the
+            // address of the export itself and not what it holds. Every other symbol here is a CFTypeRef
+            // global and needs the dereference; these two do not, and mixing them up produces a
+            // dictionary that does not retain its contents.
+            var keyCallBacks = NativeLibrary.TryGetExport(cf, "kCFTypeDictionaryKeyCallBacks", out var k)
+                ? k
+                : IntPtr.Zero;
+
+            var valueCallBacks = NativeLibrary.TryGetExport(cf, "kCFTypeDictionaryValueCallBacks", out var v)
+                ? v
+                : IntPtr.Zero;
+
+            return new MacSymbols
+            {
+                True = MacSecurity.Constant(cf, "kCFBooleanTrue"),
+                TypeDictionaryKeyCallBacks = keyCallBacks,
+                TypeDictionaryValueCallBacks = valueCallBacks,
+
+                Class = MacSecurity.Constant(sec, "kSecClass"),
+                ClassKey = MacSecurity.Constant(sec, "kSecClassKey"),
+                ReturnRef = MacSecurity.Constant(sec, "kSecReturnRef"),
+                UseDataProtectionKeychain = MacSecurity.Constant(sec, "kSecUseDataProtectionKeychain"),
+                UseOperationPrompt = MacSecurity.Constant(sec, "kSecUseOperationPrompt"),
+
+                AttrKeyType = MacSecurity.Constant(sec, "kSecAttrKeyType"),
+                AttrKeySizeInBits = MacSecurity.Constant(sec, "kSecAttrKeySizeInBits"),
+                AttrTokenId = MacSecurity.Constant(sec, "kSecAttrTokenID"),
+                AttrIsPermanent = MacSecurity.Constant(sec, "kSecAttrIsPermanent"),
+                AttrApplicationTag = MacSecurity.Constant(sec, "kSecAttrApplicationTag"),
+                AttrAccessControl = MacSecurity.Constant(sec, "kSecAttrAccessControl"),
+                PrivateKeyAttrs = MacSecurity.Constant(sec, "kSecPrivateKeyAttrs"),
+
+                KeyTypeEcSecPrimeRandom = MacSecurity.Constant(sec, "kSecAttrKeyTypeECSECPrimeRandom"),
+                TokenIdSecureEnclave = MacSecurity.Constant(sec, "kSecAttrTokenIDSecureEnclave"),
+                AccessibleWhenUnlockedThisDeviceOnly =
+                    MacSecurity.Constant(sec, "kSecAttrAccessibleWhenUnlockedThisDeviceOnly"),
+
+                EciesAlgorithm = MacSecurity.Constant(
+                    sec,
+                    "kSecKeyAlgorithmECIESEncryptionCofactorX963SHA256AESGCM"),
+            };
+        }
+        catch (Exception exception) when (exception is DllNotFoundException or BadImageFormatException)
+        {
+            return new MacSymbols();
+        }
+    }
+
+    private bool AllResolved() =>
+        True != IntPtr.Zero
+        && TypeDictionaryKeyCallBacks != IntPtr.Zero
+        && TypeDictionaryValueCallBacks != IntPtr.Zero
+        && Class != IntPtr.Zero
+        && ClassKey != IntPtr.Zero
+        && ReturnRef != IntPtr.Zero
+        && UseDataProtectionKeychain != IntPtr.Zero
+        && UseOperationPrompt != IntPtr.Zero
+        && AttrKeyType != IntPtr.Zero
+        && AttrKeySizeInBits != IntPtr.Zero
+        && AttrTokenId != IntPtr.Zero
+        && AttrIsPermanent != IntPtr.Zero
+        && AttrApplicationTag != IntPtr.Zero
+        && AttrAccessControl != IntPtr.Zero
+        && PrivateKeyAttrs != IntPtr.Zero
+        && KeyTypeEcSecPrimeRandom != IntPtr.Zero
+        && TokenIdSecureEnclave != IntPtr.Zero
+        && AccessibleWhenUnlockedThisDeviceOnly != IntPtr.Zero
+        && EciesAlgorithm != IntPtr.Zero;
+}

src/DodoSSH.Client.App/Platform/RpmVersion.cs

@@ -0,0 +1,213 @@
+namespace DodoSSH.Client.App.Platform;
+
+/// <summary>
+/// Compares two RPM version strings the way <c>rpm</c> itself does.
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>Why this exists rather than a SemVer comparison.</b> <c>AndroidUpdateChannel</c> settles the same
+/// question and its reasoning carries over exactly: the comparison has to be the one the installer will
+/// make, because anything else can offer an update the platform then refuses. On the phone that number is
+/// Android's <c>versionCode</c>; here it is the RPM version, and the authority that reads it is
+/// <c>rpm</c> — so this is <c>rpmvercmp</c> and not an approximation of it.
+/// </para>
+/// <para>
+/// It matters in a specific direction. The nightly channel is rolling and CI replaces it on every push, so
+/// the published version normally only rises — but a revert, a force-push or a re-run of an older commit
+/// can lower it, and a client comparing "is this different from mine" would then offer a downgrade that
+/// PackageKit declines. The user sees an update that fails, every six hours, forever.
+/// </para>
+/// <para>
+/// <b>The algorithm is rpm's, including the parts that look like bugs.</b> Segments of digits and segments
+/// of letters are compared separately and never against each other; everything that is neither is a
+/// separator, so <c>1.0_1</c> and <c>1.0.1</c> are equal. Leading zeros are stripped, so <c>1.01</c> and
+/// <c>1.1</c> are equal too. <c>~</c> sorts before anything including the end of the string, which is what
+/// makes <c>0.0.1~alpha.0.143</c> older than <c>0.0.1</c> and is the whole reason
+/// <c>scripts/ci-linux-rpm.sh</c> translates a SemVer prerelease to one. <c>^</c> sorts after everything
+/// except the end of the string.
+/// </para>
+/// <para>
+/// Verified against <c>rpm</c>'s own answers rather than against a reading of the source: the cases in
+/// <c>RpmVersionTests</c> were generated by running <c>rpm.vercmp</c> in a container, so a disagreement is
+/// a disagreement with the program that will actually decide.
+/// </para>
+/// </remarks>
+internal static class RpmVersion
+{
+    /// <summary>
+    /// Orders two RPM version strings.
+    /// </summary>
+    /// <param name="left">The left version.</param>
+    /// <param name="right">The right version.</param>
+    /// <returns>-1, 0 or 1, as <c>rpmvercmp</c> would answer.</returns>
+    /// <remarks>
+    /// Null and empty are accepted and treated as equal to each other and older than anything else, which
+    /// <c>rpm</c>'s own Lua binding refuses outright. The refusal is right for a packaging tool and wrong
+    /// here: this compares a version read off a feed, and a feed that answered with an empty string should
+    /// leave the client on the build it has rather than throwing on a background timer.
+    /// </remarks>
+    internal static int Compare(string? left, string? right)
+    {
+        var one = left ?? string.Empty;
+        var two = right ?? string.Empty;
+
+        if (string.Equals(one, two, StringComparison.Ordinal))
+        {
+            return 0;
+        }
+
+        int i = 0, j = 0;
+
+        while (i < one.Length || j < two.Length)
+        {
+            // Separators. Anything that is not alphanumeric, ~ or ^ divides segments and is skipped on
+            // both sides independently — which is why 1.0_1 and 1.0.1 compare equal.
+            Skip(one, ref i);
+            Skip(two, ref j);
+
+            if (CompareMarkers(one, i, two, j) is { } marker)
+            {
+                if (marker != 0)
+                {
+                    return marker;
+                }
+
+                // Both sides carried the same marker, so step over it and keep going.
+                i++;
+                j++;
+
+                continue;
+            }
+
+            if (i >= one.Length || j >= two.Length)
+            {
+                break;
+            }
+
+            if (CompareSegments(one, ref i, two, ref j) is { } ordering)
+            {
+                return ordering;
+            }
+        }
+
+        // Both ran out together, so every segment matched and the only differences were separators.
+        if (i >= one.Length && j >= two.Length)
+        {
+            return 0;
+        }
+
+        // Whichever still has characters left has more version, and more version is newer.
+        return i < one.Length ? 1 : -1;
+    }
+
+    /// <summary>Advances past anything that divides segments rather than taking part in them.</summary>
+    private static void Skip(string value, ref int at)
+    {
+        while (at < value.Length && !IsVersionCharacter(value[at]))
+        {
+            at++;
+        }
+    }
+
+    /// <summary>
+    /// Orders the two sides when either is at a <c>~</c> or a <c>^</c>.
+    /// </summary>
+    /// <returns>
+    /// Null when neither side is at a marker, so the caller should compare segments; 0 when both are at
+    /// the same one, so the caller should step over it; otherwise the ordering.
+    /// </returns>
+    /// <remarks>
+    /// <c>~</c> sorts before everything including the end of the string, which is what carries the whole
+    /// prerelease ordering. <c>^</c> sorts after everything <em>except</em> the end of the string, so
+    /// 1.0^post is newer than 1.0 and 1.2.3 is older than 1.2.3^ — the end-of-string checks coming first
+    /// is the difference between the two halves of that sentence.
+    /// </remarks>
+    private static int? CompareMarkers(string one, int i, string two, int j)
+    {
+        var oneTilde = i < one.Length && one[i] == '~';
+        var twoTilde = j < two.Length && two[j] == '~';
+
+        if (oneTilde || twoTilde)
+        {
+            return oneTilde && twoTilde ? 0 : oneTilde ? -1 : 1;
+        }
+
+        var oneCaret = i < one.Length && one[i] == '^';
+        var twoCaret = j < two.Length && two[j] == '^';
+
+        if (!oneCaret && !twoCaret)
+        {
+            return null;
+        }
+
+        if (i >= one.Length)
+        {
+            return -1;
+        }
+
+        if (j >= two.Length)
+        {
+            return 1;
+        }
+
+        return oneCaret && twoCaret ? 0 : oneCaret ? -1 : 1;
+    }
+
+    /// <summary>
+    /// Takes one segment from each side and orders them, advancing past both.
+    /// </summary>
+    /// <returns>The ordering, or null if the two segments were equal and the walk should continue.</returns>
+    private static int? CompareSegments(string one, ref int i, string two, ref int j)
+    {
+        var numeric = char.IsAsciiDigit(one[i]);
+
+        var start1 = i;
+        var start2 = j;
+
+        Take(one, ref i, numeric);
+        Take(two, ref j, numeric);
+
+        var segment1 = one.AsSpan(start1, i - start1);
+        var segment2 = two.AsSpan(start2, j - start2);
+
+        // The two sides are at segments of different kinds — digits against letters. rpm calls the
+        // numeric one newer, always, on the reasoning that a version that has started counting is further
+        // along than one still spelling. segment1 is never empty: its kind was chosen from its own first
+        // character.
+        if (segment2.IsEmpty)
+        {
+            return numeric ? 1 : -1;
+        }
+
+        if (numeric)
+        {
+            segment1 = segment1.TrimStart('0');
+            segment2 = segment2.TrimStart('0');
+
+            // Longer, once the zeros are gone, is larger. Compared before the characters are, because a
+            // string comparison would call "9" larger than "10".
+            if (segment1.Length != segment2.Length)
+            {
+                return segment1.Length > segment2.Length ? 1 : -1;
+            }
+        }
+
+        var ordering = segment1.CompareTo(segment2, StringComparison.Ordinal);
+
+        return ordering == 0 ? null : ordering > 0 ? 1 : -1;
+    }
+
+    /// <summary>Advances over one run of digits, or one run of letters.</summary>
+    private static void Take(string value, ref int at, bool numeric)
+    {
+        while (at < value.Length
+               && (numeric ? char.IsAsciiDigit(value[at]) : char.IsAsciiLetter(value[at])))
+        {
+            at++;
+        }
+    }
+
+    /// <summary>Whether this character takes part in the comparison rather than dividing it.</summary>
+    private static bool IsVersionCharacter(char value) =>
+        char.IsAsciiLetterOrDigit(value) || value == '~' || value == '^';
+}

src/DodoSSH.Client.App/Platform/SingleInstance.cs

@@ -0,0 +1,372 @@
+using System.IO.Pipes;
+using System.Security.Cryptography;
+using System.Text;
+using DodoSSH.Client.Session;
+
+namespace DodoSSH.Client.App.Platform;
+
+/// <summary>
+/// One desktop instance per profile, and a way for a second launch to raise the first.
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>Per profile rather than per machine, which decides two things at once.</b> The lock lives in
+/// <see cref="ClientPaths.InstanceLockFile"/>, so a nightly and a release build never exclude each other —
+/// they are separate applications by design, ADR 0013 — and two people signed in to one machine each get
+/// their own instance. A global mutex would have had to re-derive both of those from somewhere.
+/// </para>
+/// <para>
+/// <b>Why one instance at all.</b> The cache is a SQLite file with an outbox in it, and the whole design
+/// assumes one writer: <see cref="ClientPaths"/> says so about two machines sharing a profile through a
+/// cloud sync client, and two processes on one machine is the same hazard with a shorter path to it. The
+/// terminal's loopback listener is the visible half — the second process fails to bind, gets a port of its
+/// own or refuses to start a workspace at all — and the cache is the half that would quietly corrupt.
+/// </para>
+/// <para>
+/// <b>A file handle rather than a named mutex</b>, because the operating system releases it when the
+/// process ends <em>however</em> it ends. A named mutex abandoned by a killed process is a state both
+/// platforms model differently and neither models simply; an exclusive handle is closed by the kernel on
+/// Windows and by the file descriptor going away on Unix, so there is no such thing as a stale lock here.
+/// </para>
+/// <para>
+/// <b>And a named pipe for the other half, because exiting silently is not an answer.</b> Somebody who
+/// double-clicks the icon and sees nothing happen has been told the application is broken. The second
+/// process asks the first to come forward and then leaves, which is what every desktop application does and
+/// what makes the exclusion invisible rather than mysterious.
+/// </para>
+/// </remarks>
+internal static class SingleInstance
+{
+    /// <summary>The one thing a second launch has to say.</summary>
+    /// <remarks>
+    /// Compared against a constant rather than parsed, for the reason <c>MainWindow</c>'s release-focus
+    /// message is: one message exists and treating anything else as a command would be a wider door than
+    /// this needs. Anything may connect to a per-user pipe — it is not a security boundary — so what arrives
+    /// is checked rather than trusted.
+    /// </remarks>
+    private const string ComeForward = "dodossh.come-forward";
+
+    /// <summary>How long a second launch waits for the first to answer its pipe.</summary>
+    /// <remarks>
+    /// Short, because the pipe is local and the process on the other end is either listening or not. It is
+    /// not a bound on anything slow; it is how long to wait before concluding that whoever holds the lock
+    /// cannot be talked to.
+    /// </remarks>
+    private const int AnswerTimeoutMilliseconds = 500;
+
+    /// <summary>How long to keep trying for a lock that is held by nobody who answers.</summary>
+    /// <remarks>
+    /// <para>
+    /// ◆ <b>This is the window an update relaunch lands in, and without it an update would look like a
+    /// crash.</b> Velopack's apply ends this process and its updater then starts the new build; the two are
+    /// ordered — the updater waits on the old process id — but the whole sequence is one where a lock held a
+    /// moment too long would make the relaunched application exit on the spot, having apparently uninstalled
+    /// itself. The same window covers an ordinary double-double-click, where the first process holds the file
+    /// and has not reached its pipe yet.
+    /// </para>
+    /// <para>
+    /// Only reached when nobody answers the pipe. A live instance replies in milliseconds and the second
+    /// launch is gone before any of this matters.
+    /// </para>
+    /// </remarks>
+    private static readonly TimeSpan DefaultPatience = TimeSpan.FromSeconds(3);
+
+    /// <summary>How long to wait between attempts within that window.</summary>
+    private static readonly TimeSpan RetryDelay = TimeSpan.FromMilliseconds(100);
+
+    /// <summary>
+    /// Claims this profile for this process, or asks whoever holds it to come forward.
+    /// </summary>
+    /// <returns>
+    /// A handle to hold for the life of the process, or <see langword="null"/> when another instance has it —
+    /// in which case that instance has already been asked to raise its window and this process should end.
+    /// </returns>
+    /// <param name="paths">The profile to claim. Its directory is created if it is not there.</param>
+    /// <param name="patience">
+    /// How long to keep trying a lock nobody answers for, or null for <see cref="DefaultPatience"/>. Named
+    /// by a test so the give-up path does not cost the suite three seconds; the application never passes it.
+    /// </param>
+    internal static IDisposable? Claim(ClientPaths paths, TimeSpan? patience = null)
+    {
+        ArgumentNullException.ThrowIfNull(paths);
+
+        try
+        {
+            paths.EnsureCreated();
+        }
+        catch (Exception exception) when (exception is IOException or UnauthorizedAccessException)
+        {
+            // No profile directory means no lock file, and every build before this one started anyway. What
+            // follows will fail on its own terms and say something better than this could.
+            return NoClaim.Instance;
+        }
+
+        // Counted rather than clocked. Thread.Sleep is not a clock and reading one here would be pretending
+        // otherwise — what this bounds is the number of attempts, and the wall time is whatever the platform
+        // makes of them.
+        var attemptsLeft = (int)Math.Ceiling(
+            (patience ?? DefaultPatience).TotalMilliseconds / RetryDelay.TotalMilliseconds);
+
+        while (true)
+        {
+            if (TryTake(paths.InstanceLockFile) is { } taken)
+            {
+                return taken;
+            }
+
+            // Asked before waiting, so the ordinary case — an instance that is up and running — costs one
+            // round trip on a local pipe rather than the whole window below.
+            if (AskTheRunningOneToComeForward(paths))
+            {
+                return null;
+            }
+
+            if (attemptsLeft-- <= 0)
+            {
+                // Held by something that will not talk. Giving up rather than starting anyway, because what
+                // is being protected is a database with one writer, and "probably nobody" is not an answer
+                // worth corrupting it for.
+                return null;
+            }
+
+            Thread.Sleep(RetryDelay);
+        }
+    }
+
+    /// <summary>
+    /// Answers the launches that were refused, for as long as the returned handle is held.
+    /// </summary>
+    /// <remarks>
+    /// <paramref name="comeForward"/> is invoked on a thread pool thread, because that is where a pipe
+    /// accept completes. A caller that touches a window has to marshal.
+    /// </remarks>
+    /// <param name="paths">The profile this instance claimed.</param>
+    /// <param name="comeForward">What to do when another launch asks for the window.</param>
+    internal static IAsyncDisposable Listen(ClientPaths paths, Action comeForward)
+    {
+        ArgumentNullException.ThrowIfNull(paths);
+        ArgumentNullException.ThrowIfNull(comeForward);
+
+        var listener = new Doorbell(PipeNameFor(paths), comeForward);
+
+        listener.Start();
+
+        return listener;
+    }
+
+    /// <summary>Tells the instance that holds this profile to raise its window.</summary>
+    /// <returns>Whether anything was listening.</returns>
+    /// <remarks>
+    /// Internal rather than private because it is the half a test can drive: the exclusion is a file handle
+    /// and proves itself, and this is the part with a protocol in it.
+    /// </remarks>
+    internal static bool AskTheRunningOneToComeForward(ClientPaths paths)
+    {
+        ArgumentNullException.ThrowIfNull(paths);
+
+        try
+        {
+            using var pipe = new NamedPipeClientStream(
+                serverName: ".",
+                PipeNameFor(paths),
+                PipeDirection.Out);
+
+            pipe.Connect(AnswerTimeoutMilliseconds);
+
+            using var writer = new StreamWriter(pipe, Encoding.UTF8) { AutoFlush = true };
+
+            writer.WriteLine(ComeForward);
+
+            return true;
+        }
+        catch (Exception exception)
+            when (exception is TimeoutException or IOException or UnauthorizedAccessException)
+        {
+            // Nothing there, or nothing willing. The caller decides what that means; here it is only ever
+            // "no".
+            return false;
+        }
+    }
+
+    /// <summary>Takes the lock, or answers null if somebody else has it.</summary>
+    /// <remarks>
+    /// The file is left in place rather than deleted on close. <c>FileOptions.DeleteOnClose</c> looks
+    /// tidier and is worse: it makes the file's existence part of the protocol, so a second process opening
+    /// it while the first is closing gets a handle to something on its way out. An empty file that is always
+    /// there and sometimes locked has one fewer state.
+    /// </remarks>
+    private static IDisposable? TryTake(string lockFile)
+    {
+        try
+        {
+            return new FileStream(
+                lockFile,
+                FileMode.OpenOrCreate,
+                FileAccess.ReadWrite,
+                FileShare.None,
+                bufferSize: 1,
+                FileOptions.None);
+        }
+        catch (IOException)
+        {
+            // Somebody else holds it. The one expected failure, and the whole point.
+            return null;
+        }
+        catch (UnauthorizedAccessException)
+        {
+            // A profile this user cannot write to, which is a broken installation rather than a second
+            // instance. Answered as a claim so the application starts and fails on something legible.
+            return NoClaim.Instance;
+        }
+    }
+
+    /// <summary>
+    /// The pipe both sides name, derived from the profile so that it is scoped exactly as the lock is.
+    /// </summary>
+    /// <remarks>
+    /// A digest of the path rather than the path itself: a pipe name may not contain a directory separator
+    /// on Windows, and on Unix the name becomes a file in a temporary directory whose length is bounded.
+    /// Eight bytes of SHA-256, which is not doing security work — it is making one string out of another
+    /// deterministically, and <c>string.GetHashCode</c> cannot because it is randomised per process.
+    /// </remarks>
+    private static string PipeNameFor(ClientPaths paths)
+    {
+        var digest = SHA256.HashData(Encoding.UTF8.GetBytes(paths.DataDirectory));
+
+        return "dodossh-" + Convert.ToHexStringLower(digest.AsSpan(0, 8));
+    }
+
+    /// <summary>A claim over nothing, for the cases where there was nothing to claim.</summary>
+    private sealed class NoClaim : IDisposable
+    {
+        internal static NoClaim Instance { get; } = new();
+
+        public void Dispose()
+        {
+            // Nothing was taken.
+        }
+    }
+
+    /// <summary>The accept loop that answers refused launches.</summary>
+    /// <remarks>
+    /// <para>
+    /// One server instance for the life of the doorbell, disconnected between callers rather than disposed
+    /// and rebuilt. The loop is what makes a third and fourth launch work as well as the second.
+    /// </para>
+    /// <para>
+    /// ◆ <b>THE ONE INSTANCE IS THE WHOLE CORRECTNESS OF THIS CLASS ON LINUX, AND A FRESH ONE PER CALLER
+    /// LOSES LAUNCHES.</b> This was written the obvious way — a <c>using</c> inside the loop, a new stream
+    /// per connection — which is fine on Windows and quietly wrong on Unix, where a named pipe is a Unix
+    /// domain socket. Disposing the server there closes the <em>listening</em> socket and unlinks its path,
+    /// so a second launch that connected while the first was being served has its connection dropped on the
+    /// floor: its <c>Connect</c> returned, its write succeeded into a buffer, and nobody ever read it. The
+    /// launch was told it had rung the doorbell and the window never came forward.
+    /// </para>
+    /// <para>
+    /// Measured rather than reasoned about, in a container, because the two platforms disagree and this
+    /// machine is only one of them: three rings back to back are answered three times with one instance and
+    /// once with a fresh one per connection. <c>Disconnect</c> ends the conversation and keeps the socket, so
+    /// there is never a moment with nothing listening.
+    /// </para>
+    /// <para>
+    /// Which makes <c>leaveOpen</c> on the reader load-bearing as well. A <see cref="StreamReader"/> disposes
+    /// the stream it was given, and the stream is now the server that has to survive into the next
+    /// iteration — so the reader that was harmless per-connection would close the doorbell after one ring.
+    /// </para>
+    /// </remarks>
+    private sealed class Doorbell : IAsyncDisposable
+    {
+        private readonly string name;
+        private readonly Action comeForward;
+        private readonly CancellationTokenSource lifetime = new();
+
+        private Task? loop;
+
+        internal Doorbell(string name, Action comeForward)
+        {
+            this.name = name;
+            this.comeForward = comeForward;
+        }
+
+        internal void Start() => loop = RunAsync(lifetime.Token);
+
+        /// <inheritdoc />
+        public async ValueTask DisposeAsync()
+        {
+            await lifetime.CancelAsync().ConfigureAwait(false);
+
+            if (loop is { } running)
+            {
+                // Awaited rather than abandoned, so the pipe name is free by the time this returns — which
+                // matters for exactly one caller: an update that is about to relaunch this application and
+                // would otherwise race its own successor for the name.
+                await running.ConfigureAwait(false);
+            }
+
+            lifetime.Dispose();
+        }
+
+        private async Task RunAsync(CancellationToken cancellationToken)
+        {
+            using var server = new NamedPipeServerStream(
+                name,
+                PipeDirection.In,
+                maxNumberOfServerInstances: 1,
+                PipeTransmissionMode.Byte,
+                PipeOptions.Asynchronous);
+
+            while (!cancellationToken.IsCancellationRequested)
+            {
+                try
+                {
+                    await server.WaitForConnectionAsync(cancellationToken).ConfigureAwait(false);
+
+                    // leaveOpen, because this server outlives the conversation. See the remark above.
+                    using (var reader = new StreamReader(server, Encoding.UTF8, false, 1024, leaveOpen: true))
+                    {
+                        var said = await reader.ReadLineAsync(cancellationToken).ConfigureAwait(false);
+
+                        if (string.Equals(said, ComeForward, StringComparison.Ordinal))
+                        {
+                            comeForward();
+                        }
+                    }
+
+                    HangUp(server);
+                }
+                catch (OperationCanceledException)
+                {
+                    return;
+                }
+                catch (IOException)
+                {
+                    // A client that connected and hung up mid-sentence. Round again: giving up here would
+                    // leave every later launch with nothing to talk to, which is the silent-exit behaviour
+                    // this class exists to avoid.
+                    HangUp(server);
+                }
+            }
+        }
+
+        /// <summary>Ends one conversation without ending the doorbell.</summary>
+        /// <remarks>
+        /// Guarded, because <c>Disconnect</c> throws on a server that is not connected — which is the state
+        /// an <see cref="IOException"/> above may already have left it in, and the state a second call would
+        /// find. Throwing out of the error path would take the loop with it.
+        /// </remarks>
+        private static void HangUp(NamedPipeServerStream server)
+        {
+            try
+            {
+                if (server.IsConnected)
+                {
+                    server.Disconnect();
+                }
+            }
+            catch (InvalidOperationException)
+            {
+                // Already disconnected. Nothing to end.
+            }
+        }
+    }
+}

src/DodoSSH.Client.App/Platform/VelopackUpdateChannel.cs

@@ -1,3 +1,4 @@
+using System.Runtime.Versioning;
 using DodoSSH.Client.Session;
 using Velopack;
 using Velopack.Sources;
@@ -31,7 +32,19 @@ internal static class UpdateChannels
     /// </remarks>
     internal static IUpdateChannel ForThisMachine()
     {
-        if (!OperatingSystem.IsWindows())
+        // ◆ Linux is answered first and separately, because it is not a Velopack platform at all.
+        //
+        // This used to be the place that said so — a negation over the two packaged platforms, with a note
+        // that a future AppImage should be an addition here rather than something that had silently
+        // half-happened. It is an addition, and it is not an AppImage: the Linux head ships as an RPM and
+        // is updated by the machine's own package manager, so its channel fetches and hands off rather
+        // than replacing anything. See LinuxUpdateChannel and ADR 0013 rule 11.
+        if (OperatingSystem.IsLinux())
+        {
+            return LinuxChannel();
+        }
+
+        if (!OperatingSystem.IsWindows() && !OperatingSystem.IsMacOS())
         {
             return new UnavailableUpdateChannel();
         }
@@ -52,17 +65,53 @@ internal static class UpdateChannels
             return new UnavailableUpdateChannel();
         }
     }
+
+    /// <summary>The Linux channel, if this copy is one a package manager put here.</summary>
+    /// <remarks>
+    /// <para>
+    /// <b>The question Velopack answers with <c>IsInstalled</c>, answered by looking at where this process
+    /// lives.</b> The RPM owns <c>/opt/DodoSSH</c> or <c>/opt/DodoSSH.Nightly</c> — the spec puts it there
+    /// and nothing else writes to <c>/opt</c> under those names — so a build running from anywhere else is
+    /// a checkout somebody started, an extracted archive, or a tree copied by hand. None of those can be
+    /// updated by installing a package over them, and offering it would be an offer that cannot be carried
+    /// out.
+    /// </para>
+    /// <para>
+    /// Checked by path rather than by asking rpm which package owns the executable, which would be the
+    /// more thorough answer and costs a process launch on every start for a distinction nothing acts on
+    /// differently. The install root is a constant in two places — here and the packaging script — and
+    /// they are named in each other's comments.
+    /// </para>
+    /// </remarks>
+    [SupportedOSPlatform("linux")]
+    private static IUpdateChannel LinuxChannel()
+    {
+        var root = DesktopChannel.IsNightly ? "/opt/DodoSSH.Nightly" : "/opt/DodoSSH";
+
+        var here = AppContext.BaseDirectory.TrimEnd(Path.DirectorySeparatorChar);
+
+        return string.Equals(here, root, StringComparison.Ordinal)
+            ? new LinuxUpdateChannel(DesktopChannel.Name)
+            : new UnavailableUpdateChannel();
+    }
 }
 
 /// <summary>
-/// The Windows update channel, backed by Velopack against the project's own forge.
+/// The desktop update channel, backed by Velopack against the project's own forge.
 /// </summary>
 /// <remarks>
 /// <para>
-/// The one file in the repository that names Velopack. It lives beside <c>WindowsDeviceKeyStore</c>
-/// rather than in a project of its own because it is the same kind of thing — a Windows-only
-/// implementation of an interface declared in <c>DodoSSH.Client.Session</c> — and because
-/// <c>DodoSSH.Client.Shell</c> is shared with the Android head, which must never acquire an updater.
+/// The one file in the repository that names Velopack. It lives beside the platform key stores rather
+/// than in a project of its own because it is the same kind of thing — a desktop-only implementation of
+/// an interface declared in <c>DodoSSH.Client.Session</c> — and because <c>DodoSSH.Client.Shell</c> is
+/// shared with the Android head, which must never acquire an updater.
+/// </para>
+/// <para>
+/// <b>One class for both desktop platforms, where the key stores are one class each.</b> The difference
+/// is where the platform knowledge sits. A key store is platform knowledge from top to bottom: different
+/// hardware, different API, different failure modes. Velopack's <c>UpdateManager</c> has already absorbed
+/// all of that, and what is left over — check, download, apply, restart — is identical on the two. The
+/// only thing that differs is which string names the feed, and that is <see cref="ChannelFor"/>.
 /// </para>
 /// <para>
 /// See <c>docs/adr/0013-desktop-distribution-and-updates.md</c>.
@@ -103,7 +152,45 @@ internal sealed class VelopackUpdateChannel : IUpdateChannel
     /// but unsaid on one side and stated on the other is how a feed goes quiet with no error anywhere:
     /// the check succeeds, finds nothing, and reports that the client is up to date forever.
     /// </remarks>
-    private const string ReleaseChannel = "win";
+    private const string WindowsReleaseChannel = "win";
+
+    /// <summary>
+    /// The nightly channel, which is a different name rather than the same one on a different tag.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// A contract with the <c>desktop nightly</c> job in <c>.github/workflows/ci.yml</c>, which passes
+    /// this word to both <c>vpk pack</c> and <c>vpk upload</c>. The name reaches the wire: Velopack
+    /// publishes its index as <c>releases.win-nightly.json</c> and looks for exactly that file, so a
+    /// disagreement between the two sides is a channel that answers nothing, forever, without an error.
+    /// </para>
+    /// <para>
+    /// <b>Two names rather than one name on two tags, and that is the part doing the work.</b> Both
+    /// channels are published to the same repository, so a client that read the other's index could be
+    /// offered a package built under a different pack id. Velopack would refuse it, but at the far end of
+    /// a download somebody watched. A channel each means neither ever sees the other's releases at all.
+    /// </para>
+    /// </remarks>
+    private const string WindowsNightlyChannel = "win-nightly";
+
+    /// <summary>The macOS release channel, and Velopack's own default there.</summary>
+    /// <remarks>
+    /// A contract with <c>scripts/release-macos.sh</c>, exactly as the Windows pair is one with the
+    /// PowerShell script. Stated for the same reason, which applies with more force here: the four
+    /// channels all publish to one repository, so the only thing keeping a Mac from being offered a
+    /// <c>win</c> package is that it never reads that index.
+    /// </remarks>
+    private const string MacReleaseChannel = "osx";
+
+    /// <summary>The macOS nightly channel.</summary>
+    /// <remarks>
+    /// Named here and not yet published by anything. The CI job for the macOS head builds and bundles
+    /// and deliberately uploads nothing — see the packaging step in <c>ci.yml</c> — so a nightly macOS
+    /// build checking this feed finds an empty channel and reports itself up to date, which is the
+    /// correct behaviour for a channel with no publisher. The name exists so that turning the publisher
+    /// on later is one job rather than a job plus a rename that has to reach every installed client.
+    /// </remarks>
+    private const string MacNightlyChannel = "osx-nightly";
 
     private readonly UpdateManager manager;
 
@@ -117,15 +204,37 @@ internal sealed class VelopackUpdateChannel : IUpdateChannel
     /// </remarks>
     private UpdateInfo? found;
 
+    /// <summary>
+    /// The last thing actually fetched, kept separately from the last thing found.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// ◆ <b>Because the two can disagree, and when they do it is an apply that pays for it.</b> A check
+    /// overwrites <see cref="found"/>, and the shell now checks while a build is already downloaded and
+    /// waiting — that is how it replaces a pending build when the feed has moved on, and how it takes one
+    /// last look on the way to a restart. Either of those can land a newer answer in <c>found</c> and then
+    /// fail to fetch it, or be given up on halfway; the shell's answer is to restart with the build it
+    /// already has, and <see cref="Matched"/> resolving only against <c>found</c> would have thrown at it.
+    /// </para>
+    /// <para>
+    /// Which turned a slow forge into a refused restart: the offer on screen was real, the package on disk
+    /// was real, and the only thing wrong was that this class had forgotten which one it was.
+    /// </para>
+    /// </remarks>
+    private UpdateInfo? fetched;
+
     internal VelopackUpdateChannel(UpdateManager manager) => this.manager = manager;
 
     /// <inheritdoc />
     public bool IsSupported => true;
 
-    /// <summary>Always, on this head.</summary>
+    /// <summary>Always, on this head, on either platform.</summary>
     /// <remarks>
-    /// Velopack's apply runs <c>Update.exe</c> over this installation and restarts it, so the process is
-    /// gone by the time anything could have asked a question. The phone's is the other answer; see
+    /// Velopack's apply hands off to a separate updater process — <c>Update.exe</c> on Windows, the
+    /// <c>UpdateMac</c> helper inside the bundle on macOS — which replaces this installation and
+    /// relaunches it, so the process is gone by the time anything could have asked a question. The
+    /// mechanism differs and the answer does not, which is why this is a constant rather than another
+    /// thing <see cref="ChannelFor"/> would have to decide. The phone's is the other answer; see
     /// <see cref="IUpdateChannel.ApplyingEndsTheProcess"/> for what the caller does differently.
     /// </remarks>
     public bool ApplyingEndsTheProcess => true;
@@ -138,10 +247,63 @@ internal sealed class VelopackUpdateChannel : IUpdateChannel
     /// </remarks>
     public string CurrentVersion => ClientVersion.Current;
 
-    internal static UpdateManager CreateManager() =>
-        new(
-            new GiteaSource(RepositoryUrl, accessToken: null, prerelease: false),
-            new UpdateOptions { ExplicitChannel = ReleaseChannel });
+    /// <summary>
+    /// The updater for this build's channel.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// <b>The prerelease flag is the half Velopack cannot work out for itself.</b> The channel could be
+    /// left to the installed layout — Velopack records what a package was built with — but whether to
+    /// consider prereleases is a property of the feed rather than of the install, and it decides more
+    /// than it looks. The nightly is published as a prerelease deliberately: the Android head's release
+    /// channel reads <c>releases/latest</c>, which skips prereleases, so a desktop nightly published as a
+    /// stable release would become the newest release in this repository and the phone's release channel
+    /// would start finding no Android manifest on it. One flag here keeps the two heads out of each
+    /// other's way.
+    /// </para>
+    /// <para>
+    /// The release channel takes <c>false</c>, so it cannot see the nightly at all — which is the property
+    /// that matters most, because that is the direction where a mistake would put an unsigned CI build on
+    /// a machine somebody trusts with their credentials.
+    /// </para>
+    /// </remarks>
+    internal static UpdateManager CreateManager()
+    {
+        var nightly = DesktopChannel.IsNightly;
+
+        return new UpdateManager(
+            new GiteaSource(RepositoryUrl, accessToken: null, prerelease: nightly),
+            new UpdateOptions { ExplicitChannel = ChannelFor(nightly) });
+    }
+
+    /// <summary>
+    /// The one of the four channel names this build belongs to.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// Two independent axes — which platform, and which of that platform's two channels — and they are
+    /// resolved in one place so that neither can be answered differently somewhere else. The platform
+    /// half is the running OS rather than anything recorded in the build, because a package can only
+    /// ever be applied on the platform it was built for; the channel half comes from assembly metadata,
+    /// because a release build and a nightly are the same bytes on the same OS and only the metadata
+    /// tells them apart.
+    /// </para>
+    /// <para>
+    /// Windows is the fallback rather than a third branch. Only Windows and macOS reach here at all —
+    /// <see cref="UpdateChannels.ForThisMachine"/> is the gate — so the alternative would be an
+    /// unreachable throw, and an unreachable throw in the middle of the updater is a thing somebody
+    /// later has to reason about to discover it cannot happen.
+    /// </para>
+    /// </remarks>
+    private static string ChannelFor(bool nightly)
+    {
+        if (OperatingSystem.IsMacOS())
+        {
+            return nightly ? MacNightlyChannel : MacReleaseChannel;
+        }
+
+        return nightly ? WindowsNightlyChannel : WindowsReleaseChannel;
+    }
 
     /// <inheritdoc />
     public async Task<AvailableUpdate?> CheckAsync(CancellationToken cancellationToken)
@@ -178,7 +340,23 @@ internal sealed class VelopackUpdateChannel : IUpdateChannel
 
         // Velopack reports progress as an Action<int> and the rest of this codebase speaks IProgress<T>,
         // so the adaptation happens here rather than leaking the older shape into the view models.
-        return manager.DownloadUpdatesAsync(Matched(update), progress.Report, cancellationToken);
+        return DownloadAndRememberAsync(Matched(update), progress, cancellationToken);
+    }
+
+    /// <remarks>
+    /// Recorded after the await and not before, so that a download which threw leaves the previous fetch as
+    /// the one this channel can still apply. See <see cref="fetched"/>.
+    /// </remarks>
+    private async Task DownloadAndRememberAsync(
+        UpdateInfo info,
+        IProgress<int> progress,
+        CancellationToken cancellationToken)
+    {
+        await manager
+            .DownloadUpdatesAsync(info, progress.Report, cancellationToken)
+            .ConfigureAwait(false);
+
+        fetched = info;
     }
 
     /// <inheritdoc />
@@ -192,21 +370,39 @@ internal sealed class VelopackUpdateChannel : IUpdateChannel
     }
 
     /// <remarks>
+    /// <para>
     /// The guard exists because the seam narrows <c>UpdateInfo</c> down to a version string, so nothing in
     /// the type system stops a caller inventing one. Every legitimate caller passes back exactly what
-    /// <see cref="CheckAsync"/> returned; a mismatch is a bug in this application rather than anything a
-    /// user did, which is why it throws rather than resolving to some safe-looking default.
+    /// <see cref="CheckAsync"/> returned, or what a download of that then fetched; a mismatch is a bug in
+    /// this application rather than anything a user did, which is why it throws rather than resolving to
+    /// some safe-looking default.
+    /// </para>
+    /// <para>
+    /// What was fetched is tried before what was found, because that is the one an apply names and the one a
+    /// later check cannot take away. See <see cref="fetched"/> for how the two come to disagree.
+    /// </para>
     /// </remarks>
     private UpdateInfo Matched(AvailableUpdate update)
     {
-        if (found is not { } info
-            || !string.Equals(info.TargetFullRelease.Version.ToString(), update.Version, StringComparison.Ordinal))
+        if (Names(fetched, update))
         {
-            throw new InvalidOperationException(
-                $"No update matching {update.Version} has been found by this channel. "
-                + "Call CheckAsync and pass back what it returned.");
+            return fetched!;
         }
 
-        return info;
+        if (Names(found, update))
+        {
+            return found!;
+        }
+
+        throw new InvalidOperationException(
+            $"No update matching {update.Version} has been found or fetched by this channel. "
+            + "Call CheckAsync and pass back what it returned.");
     }
+
+    private static bool Names(UpdateInfo? info, AvailableUpdate update) =>
+        info is not null
+        && string.Equals(
+            info.TargetFullRelease.Version.ToString(),
+            update.Version,
+            StringComparison.Ordinal);
 }

src/DodoSSH.Client.App/Platform/WindowsDeviceKeyStore.cs

@@ -9,9 +9,26 @@ namespace DodoSSH.Client.App.Platform;
 /// </summary>
 /// <remarks>
 /// <para>
-/// One place decides, so nothing above has to carry a platform guard. A machine with no TPM, or one that
-/// is not Windows, gets <see cref="UnavailableDeviceKeyStore"/> and therefore keeps asking for the
-/// passphrase — which is the honest answer rather than a degraded one.
+/// One place decides, so nothing above has to carry a platform guard. A machine with no secure hardware
+/// gets <see cref="UnavailableDeviceKeyStore"/> and therefore keeps asking for the passphrase — which is
+/// the honest answer rather than a degraded one.
+/// </para>
+/// <para>
+/// <b>Every real store is asked whether it works rather than told that it does.</b> Each
+/// <c>IsSupported</c> probes by doing the thing — creating a throwaway key and discarding it — because on
+/// all three platforms the provider is present and reports itself present on machines where creating a key
+/// fails: a Windows box with no usable TPM, a Mac with no Secure Enclave, on macOS also every unsigned
+/// development build since enclave keys need a signing identity, and on Linux the ordinary case of a
+/// working TPM whose device node this user may not open. Inferring from the OS would mean each of those
+/// discovering the truth at the moment somebody tried to unlock.
+/// </para>
+/// <para>
+/// <b>The three are not equally strong, and the ordering here does not imply they are.</b> Windows and
+/// macOS put the consent gesture inside the platform — a secure desktop and a system Touch ID sheet — so
+/// the process asking never sees the secret. The Linux store's PIN is typed into a window this
+/// application draws, and only the <em>checking</em> is the TPM's. That is a real difference, it is
+/// recorded in ADR 0007 rather than smoothed over, and it is why <see cref="LinuxDeviceKeyStore"/> is
+/// tried last rather than first among platforms that could in principle overlap.
 /// </para>
 /// <para>
 /// <b>"Desktop", because the choice belongs to a head rather than to the session layer.</b> This file used
@@ -25,13 +42,37 @@ namespace DodoSSH.Client.App.Platform;
 public static class DesktopDeviceKeyStores
 {
     /// <summary>The best store this machine supports.</summary>
-    public static IDeviceKeyStore ForThisMachine(ClientPaths paths)
+    /// <param name="paths">Where a stored key lives.</param>
+    /// <param name="askForPin">
+    /// Shows the Linux PIN prompt, and is never called on the other two platforms — there the platform
+    /// raises its own dialog and this process is not in the conversation. A caller with no way to draw one
+    /// may pass a delegate answering null, which makes the Linux store decline every unlock; that is the
+    /// right shape for a headless host and the wrong one for the application, which supplies a real dialog
+    /// in <c>App.axaml.cs</c>.
+    /// </param>
+    public static IDeviceKeyStore ForThisMachine(
+        ClientPaths paths,
+        Func<DevicePinPurpose, CancellationToken, ValueTask<string?>> askForPin)
     {
         ArgumentNullException.ThrowIfNull(paths);
+        ArgumentNullException.ThrowIfNull(askForPin);
 
-        return OperatingSystem.IsWindows() && WindowsDeviceKeyStore.IsSupported()
-            ? new WindowsDeviceKeyStore(paths)
-            : new UnavailableDeviceKeyStore();
+        if (OperatingSystem.IsWindows() && WindowsDeviceKeyStore.IsSupported())
+        {
+            return new WindowsDeviceKeyStore(paths);
+        }
+
+        if (OperatingSystem.IsMacOS() && MacDeviceKeyStore.IsSupported())
+        {
+            return new MacDeviceKeyStore(paths);
+        }
+
+        if (OperatingSystem.IsLinux() && LinuxDeviceKeyStore.IsSupported())
+        {
+            return new LinuxDeviceKeyStore(paths, askForPin);
+        }
+
+        return new UnavailableDeviceKeyStore();
     }
 }
 

src/DodoSSH.Client.App/Program.cs

@@ -1,5 +1,6 @@
 using Avalonia;
 using Avalonia.Media;
+using DodoSSH.Client.App.Platform;
 using DodoSSH.Client.Session;
 using Velopack;
 
@@ -36,6 +37,27 @@ internal static class Program
 
         KeepTheWebViewProfileOutOfTheInstallDirectory();
 
+        /*
+          One instance per profile, and nothing above this line is affected by it: the installer's own runs
+          have already exited inside VelopackApp.Run, and the WebView2 environment variable is a property of
+          this process rather than of the application.
+
+          Claimed here rather than after Avalonia is configured, so that a second launch starts nothing — no
+          toolkit, no window, no splash flashing on screen before it thinks better of it. What it does
+          instead is ask the running instance to come forward, which is why this is not simply an exit.
+
+          The handle is held for as long as Main runs, which is as long as the message loop does. See
+          SingleInstance for why it is a file rather than a mutex, and for what the profile scoping buys.
+        */
+        var paths = ClientPaths.ForChannel(DesktopChannel.Name);
+
+        using var claim = SingleInstance.Claim(paths);
+
+        if (claim is null)
+        {
+            return;
+        }
+
         BuildAvaloniaApp().StartWithClassicDesktopLifetime(args);
     }
 
@@ -69,7 +91,12 @@ internal static class Program
             return;
         }
 
-        var folder = Path.Combine(ClientPaths.Default.DataDirectory, "WebView2");
+        // The same profile directory the rest of the application resolves, channel and all — a nightly
+        // pointing WebView2 at the release build's profile would put two browser profiles in one folder
+        // and hand the pair of them whichever process opened first.
+        var folder = Path.Combine(
+            ClientPaths.ForChannel(DesktopChannel.Name).DataDirectory,
+            "WebView2");
 
         try
         {
@@ -87,17 +114,25 @@ internal static class Program
 
     /// <summary>Used by the designer as well as by <see cref="Main"/>.</summary>
     /// <remarks>
-    /// <c>WithInterFont</c> registers Inter; it does not make it the default, and until this line the
-    /// application shipped a font it then declined to use — falling back to Segoe UI on Windows and to
-    /// whatever fontconfig offered on Linux. Almost nothing visible moves, because App.axaml sets
-    /// <c>MonoFont</c> on essentially every control that draws text, but the fallback behind those is now
-    /// a font that travels with the build rather than one the machine is assumed to have. The layout
-    /// suite pins the same family, and has to: see HeadlessApp.
+    /// <c>WithInterFont</c> still registers Inter, and it is still what the layout suite pins: see
+    /// HeadlessApp. What changed is which face answers when nothing more specific is asked for. The v5
+    /// design specifies Montserrat, and unlike the substitution this comment used to explain, Montserrat
+    /// now ships embedded in the shell project rather than being requested by a name the machine may not
+    /// have — see DodoSSH.Client.Shell/Assets/Fonts and its csproj entry. Inter stays registered and is
+    /// named as a fallback, so a glyph Montserrat does not cover still resolves to a font that travels
+    /// with the build rather than to whatever the platform's default sans happens to be.
     /// </remarks>
     public static AppBuilder BuildAvaloniaApp() =>
         AppBuilder.Configure<DodoSshApp>()
             .UsePlatformDetect()
             .WithInterFont()
-            .With(new FontManagerOptions { DefaultFamilyName = "avares://Avalonia.Fonts.Inter/Assets#Inter" })
+            .With(new FontManagerOptions
+            {
+                DefaultFamilyName = "avares://DodoSSH.Client.Shell/Assets/Fonts#Montserrat",
+                FontFallbacks =
+                [
+                    new FontFallback { FontFamily = new FontFamily("avares://Avalonia.Fonts.Inter/Assets#Inter") },
+                ],
+            })
             .LogToTrace();
 }

src/DodoSSH.Client.App/Views/ConnectingCard.axaml

@@ -15,6 +15,28 @@
     Showing the last terminal's pane would be a lie, and showing nothing reads as the application having
     broken, so this says which machine, as whom, and how far along it is.
 
+    ── The step list, and why it is amber ───────────────────────────────────────────────────────────────
+    "How far along it is" used to be one line of prose that never changed after the tab was created, which
+    made every slow connection look exactly like every hung one. It is now the five steps of actually
+    getting there, each lit at the moment the handshake reports it — see SshConnectionPhase, which names
+    only the boundaries a client can genuinely observe. A connection that stops therefore stops on a named
+    row, and "the host key is being checked" stops being the same screen as "the host is not answering".
+
+    Amber for the step in flight, and that is the palette's rule rather than an exception to it. Green is
+    what is true and purple is what you can press; a step still happening is neither, and it is precisely
+    the caveat-worth-reading amber exists for — see the remark above Warn in Palette.axaml. Steps behind it
+    go green as they become true, and the one a refusal landed on goes red. Nothing on the list is drawn in
+    the accent, because there is nothing on it to press.
+
+    Nothing here animates, which is the argument the transfer strip makes for its own track in
+    TransfersScreen.axaml, arriving at a screen with more reason to want a spinner. A spinner is furniture
+    invented to fill a state nobody measured; these steps are measured, so the track fills to what has
+    actually finished and then waits there. Waiting is what waiting looks like.
+
+    The list is drawn for both states rather than once per state. A refused connection has the same five
+    rows and the same track — the difference is only that one row is red and the track stops — and drawing
+    it twice would be two templates to keep identical for the sake of a colour.
+
     It obeys the occlusion rule the whole window obeys: this is Avalonia-drawn content in the WebView's own
     rectangle, so the shell collapses the terminal while it is up. IsTerminalShowing and IsConnectingShowing
     are exclusive by construction — a selected tab either has a session or it does not — which is what makes
@@ -24,29 +46,138 @@
     can be laid out by a test: WebView2's adapter refuses the headless session's thread.
   -->
 
+  <UserControl.Styles>
+
+    <!--
+      A rule per lit state over one quiet default, so that the pending weight is stated once and each state
+      that differs from it is the one line that says how.
+    -->
+    <Style Selector="TextBlock.stepcaption">
+      <Setter Property="Foreground" Value="{StaticResource TextFaint}" />
+      <Setter Property="FontSize" Value="12" />
+      <Setter Property="VerticalAlignment" Value="Center" />
+    </Style>
+    <Style Selector="TextBlock.stepcaption.done">
+      <Setter Property="Foreground" Value="{StaticResource TextDim}" />
+    </Style>
+    <Style Selector="TextBlock.stepcaption.running">
+      <Setter Property="Foreground" Value="{StaticResource WarnText}" />
+      <Setter Property="FontWeight" Value="Medium" />
+    </Style>
+    <Style Selector="TextBlock.stepcaption.stopped">
+      <Setter Property="Foreground" Value="{StaticResource DangerText}" />
+    </Style>
+
+    <!--
+      The marker beside each caption. Fixed width and centred, because four different characters on a
+      ragged left edge is a list that looks broken; see ConnectionStepViewModel.Mark for which they are.
+    -->
+    <Style Selector="TextBlock.stepmark">
+      <Setter Property="Foreground" Value="{StaticResource BorderMid}" />
+      <Setter Property="FontSize" Value="12" />
+      <Setter Property="Width" Value="14" />
+      <Setter Property="TextAlignment" Value="Center" />
+      <Setter Property="VerticalAlignment" Value="Center" />
+    </Style>
+    <Style Selector="TextBlock.stepmark.done">
+      <Setter Property="Foreground" Value="{StaticResource Live}" />
+    </Style>
+    <Style Selector="TextBlock.stepmark.running">
+      <Setter Property="Foreground" Value="{StaticResource Warn}" />
+    </Style>
+    <Style Selector="TextBlock.stepmark.stopped">
+      <Setter Property="Foreground" Value="{StaticResource Danger}" />
+    </Style>
+
+    <!--
+      The track over the list. Amber while the attempt is alive and red once it is not, so that the bar says
+      the same thing as the row it stopped on rather than staying the colour of something still being waited
+      for. Chip underneath, matching the transfer strip's track.
+    -->
+    <Style Selector="ProgressBar.steptrack">
+      <Setter Property="Height" Value="5" />
+      <Setter Property="MinHeight" Value="5" />
+      <Setter Property="CornerRadius" Value="3" />
+      <Setter Property="Background" Value="{StaticResource Chip}" />
+      <Setter Property="Foreground" Value="{StaticResource Warn}" />
+    </Style>
+    <Style Selector="ProgressBar.steptrack.stopped">
+      <Setter Property="Foreground" Value="{StaticResource Danger}" />
+    </Style>
+
+  </UserControl.Styles>
+
   <Panel>
-    <StackPanel VerticalAlignment="Center" HorizontalAlignment="Center" Spacing="14" MaxWidth="460"
+    <StackPanel VerticalAlignment="Center" HorizontalAlignment="Center" Spacing="18" MaxWidth="460"
                 Margin="24">
 
+      <!--
+        Which machine, and at what address. Both selectable, for the reason the session column's own address
+        is — see SessionSidebar.axaml: a TextBlock in Avalonia is a drawing of a string rather than a string,
+        and this pair is the one place the shell says what it is dialling *before* anything has answered.
+        That is exactly when the value is wanted somewhere else: in a ticket, in a message to whoever owns
+        the host, in the ssh command being written in the window beside this one. It is also the moment it is
+        least available anywhere else, because there is no prompt yet and may never be one — a refused
+        connection leaves this card holding the only copy on screen.
+
+        Both rather than only the name. They are two lines of one identifier, and a pair where the first can
+        be taken and the second cannot is a worse answer than either.
+
+        The trimming stays, and the tooltips are what make it safe rather than a trap. A drag reaches only
+        the glyphs that were drawn, so a name or an address wearing an ellipsis has to be taken with
+        select-all — Ctrl+A, then Ctrl+C — which takes the whole value, because the ellipsis is a drawing and
+        not a shorter string. The hover says what the whole value is, so nobody copies half of one without
+        having been told there is a half. Same argument, same shape, as the sidebar's.
+      -->
       <StackPanel Spacing="6" HorizontalAlignment="Center">
-        <TextBlock Classes="mono" Text="{Binding SelectedTab.Label}" FontSize="16" FontWeight="Medium"
-                   Foreground="{StaticResource Text}" HorizontalAlignment="Center"
-                   TextTrimming="CharacterEllipsis" />
-        <TextBlock Classes="mono" Text="{Binding SelectedTab.Address}" FontSize="11.5"
-                   Foreground="{StaticResource TextFaint}" HorizontalAlignment="Center"
-                   TextTrimming="CharacterEllipsis" />
+        <SelectableTextBlock Classes="mono" Text="{Binding SelectedTab.Label}" FontSize="16"
+                             FontWeight="Medium" Foreground="{StaticResource Text}"
+                             HorizontalAlignment="Center" TextTrimming="CharacterEllipsis"
+                             ToolTip.Tip="{Binding SelectedTab.Label}" />
+        <SelectableTextBlock Classes="mono" Text="{Binding SelectedTab.Address}" FontSize="11.5"
+                             Foreground="{StaticResource TextFaint}" HorizontalAlignment="Center"
+                             TextTrimming="CharacterEllipsis"
+                             ToolTip.Tip="{Binding SelectedTab.Address}" />
       </StackPanel>
 
       <!--
-        Two states, deliberately different. Waiting is an accent line under the host's name; a refusal is
-        the reason, in the palette's red, because it is the only place the reason will be after the user
-        navigates away from the screen that started the connection.
+        Bound to StepsDone against StepCount rather than to a percentage: five steps and a maximum of five
+        means the bar is a count of things that really finished, and the arithmetic that would turn it into
+        a percentage is exactly the arithmetic that would start inventing one.
       -->
-      <TextBlock Classes="mono" Text="{Binding SelectedTab.Status}" FontSize="12"
-                 Foreground="{StaticResource Accent}" HorizontalAlignment="Center"
-                 TextWrapping="Wrap" TextAlignment="Center"
-                 IsVisible="{Binding SelectedTab.IsConnecting}" />
+      <ProgressBar Classes="steptrack" Classes.stopped="{Binding SelectedTab.IsFailed}"
+                   Minimum="0" Maximum="{Binding SelectedTab.StepCount}"
+                   Value="{Binding SelectedTab.StepsDone, Mode=OneWay}" />
 
+      <ItemsControl ItemsSource="{Binding SelectedTab.Steps}" HorizontalAlignment="Center">
+        <ItemsControl.ItemsPanel>
+          <ItemsPanelTemplate>
+            <StackPanel Spacing="7" />
+          </ItemsPanelTemplate>
+        </ItemsControl.ItemsPanel>
+        <ItemsControl.ItemTemplate>
+          <DataTemplate x:DataType="vm:ConnectionStepViewModel">
+            <StackPanel Orientation="Horizontal" Spacing="10">
+              <TextBlock Classes="stepmark"
+                         Classes.done="{Binding IsDone}"
+                         Classes.running="{Binding IsRunning}"
+                         Classes.stopped="{Binding IsStopped}"
+                         Text="{Binding Mark}" />
+              <TextBlock Classes="stepcaption mono"
+                         Classes.done="{Binding IsDone}"
+                         Classes.running="{Binding IsRunning}"
+                         Classes.stopped="{Binding IsStopped}"
+                         Text="{Binding Caption}" />
+            </StackPanel>
+          </DataTemplate>
+        </ItemsControl.ItemTemplate>
+      </ItemsControl>
+
+      <!--
+        A refusal is the reason, in the palette's red, because it is the only place the reason will be after
+        the user navigates away from the screen that started the connection. It sits under the list rather
+        than replacing it: which row it stopped on is half the answer and the sentence is the other half.
+      -->
       <SelectableTextBlock Text="{Binding SelectedTab.Status}" FontSize="13"
                            Foreground="{StaticResource Danger}" HorizontalAlignment="Center"
                            TextWrapping="Wrap" TextAlignment="Center"
@@ -62,14 +193,36 @@
         handshake that finishes afterwards is adopted rather than dropped — see
         MainWindowViewModel.CloseTabAsync. Two buttons rather than one with a converted label, because the
         two are different decisions and only one of them abandons something still running.
-      -->
-      <Button Classes="ghost" HorizontalAlignment="Center" Content="GIVE UP"
-              Command="{Binding CloseTabCommand}" CommandParameter="{Binding SelectedTab}"
-              IsVisible="{Binding SelectedTab.IsConnecting}" />
 
-      <Button Classes="ghost" HorizontalAlignment="Center" Content="CLOSE TAB"
-              Command="{Binding CloseTabCommand}" CommandParameter="{Binding SelectedTab}"
-              IsVisible="{Binding SelectedTab.IsFailed}" />
+        Beside each, the logs. The step list is this attempt and the log is every other one, which is the
+        question both a connection taking too long and a connection just refused actually raise — has this
+        machine ever worked. It is the ordinary rail destination reached the ordinary way rather than a
+        second log grown inside this card, and leaving by it does not abandon the handshake: the tab stays
+        in the strip and the card is still here on the way back.
+      -->
+      <StackPanel Orientation="Horizontal" Spacing="10" HorizontalAlignment="Center">
+        <!--
+          TRY AGAIN rather than RECONNECT, because nothing was ever connected: this tab is a handshake that
+          was refused. It runs the same command the disconnect strip's RECONNECT does — the vault's own
+          connect path, into this same tab — and it is worth having here for the reason the commonest
+          refusal is worth retrying at all: a host that was not up yet is a host that will be in a minute.
+
+          Disabled rather than hidden when there is nothing to dial, exactly as the strip's is. See
+          MainWindowViewModel.CanReconnectSelectedTab.
+        -->
+        <Button Classes="ghost" Content="TRY AGAIN" Command="{Binding ReconnectTabCommand}"
+                CommandParameter="{Binding SelectedTab}"
+                IsEnabled="{Binding CanReconnectSelectedTab, FallbackValue=False}"
+                IsVisible="{Binding SelectedTab.IsFailed}" />
+        <Button Classes="ghost" Content="SHOW LOGS" Command="{Binding ShowScreenCommand}"
+                CommandParameter="{x:Static vm:ShellScreen.Logs}" />
+        <Button Classes="ghost" Content="GIVE UP" Command="{Binding CloseTabCommand}"
+                CommandParameter="{Binding SelectedTab}"
+                IsVisible="{Binding SelectedTab.IsConnecting}" />
+        <Button Classes="ghost" Content="CLOSE TAB" Command="{Binding CloseTabCommand}"
+                CommandParameter="{Binding SelectedTab}"
+                IsVisible="{Binding SelectedTab.IsFailed}" />
+      </StackPanel>
 
     </StackPanel>
   </Panel>

src/DodoSSH.Client.App/Views/DevicePinWindow.axaml

@@ -0,0 +1,62 @@
+<Window xmlns="https://github.com/avaloniaui"
+        xmlns:x="http://schemas.microsoft.com/winfx/2006/xaml"
+        x:Class="DodoSSH.Client.App.Views.DevicePinWindow"
+        Width="420" SizeToContent="Height"
+        CanResize="False" ShowInTaskbar="False"
+        WindowStartupLocation="CenterOwner"
+        Background="{StaticResource Surface}"
+        Title="DodoSSH">
+
+  <!--
+    The PIN that unseals this machine's device key, on Linux and nowhere else.
+
+    ◆ A WINDOW, WHERE EVERY OTHER QUESTION THIS APPLICATION ASKS IS A CARD.
+
+    ConfirmDeleteCard, HostKeyCard and the rest are overlays bound to view-model state, and that is the
+    right shape for a question about the vault: the shell owns the moment and the Android head shows the
+    same one. This is not that kind of question. It belongs to a platform integration — LinuxDeviceKeyStore
+    — and the shell it would otherwise live in is shared with the phone, which has a fingerprint and must
+    never grow a PIN box. Keeping it here keeps DodoSSH.Client.Shell platform-neutral, which is the
+    property IDeviceKeyStore exists to protect.
+
+    ◆ AND IT IS DELIBERATELY PLAIN.
+
+    Windows and macOS draw their equivalent themselves — a secure desktop and a system sheet — and this one
+    cannot be either. Dressing it up to look like a system dialog would be teaching people that a
+    convincing DodoSSH PIN box is a system dialog, which is the phishing this platform's design cannot
+    prevent and should not rehearse. It says what it is, it names the application, and it stops there.
+    ADR 0007 records the difference rather than smoothing it over.
+  -->
+
+  <StackPanel Margin="20" Spacing="12">
+
+    <TextBlock Classes="heading" FontSize="15" x:Name="Heading" TextWrapping="Wrap" />
+
+    <TextBlock Foreground="{StaticResource MutedText}" FontSize="12" TextWrapping="Wrap"
+               x:Name="Explanation" />
+
+    <!--
+      RevealPassword is off and there is no eye button. A PIN is short enough to retype and this box is
+      open on a screen somebody may be sharing.
+    -->
+    <TextBox x:Name="Pin" PasswordChar="•" PlaceholderText="PIN" FontSize="14" />
+
+    <!--
+      Only when choosing. A mistyped PIN at registration is not recoverable in any useful sense: the key is
+      sealed to it, and the only way back is to withdraw the device and register again. Asking twice costs
+      one box.
+    -->
+    <TextBox x:Name="Confirm" PasswordChar="•" PlaceholderText="PIN again" FontSize="14"
+             IsVisible="False" />
+
+    <TextBlock Foreground="{StaticResource WarnText}" FontSize="12" TextWrapping="Wrap"
+               x:Name="Problem" IsVisible="False" />
+
+    <StackPanel Orientation="Horizontal" Spacing="8" HorizontalAlignment="Right">
+      <Button x:Name="Cancel" Content="CANCEL" />
+      <Button x:Name="Accept" Classes="accent" Content="UNLOCK" />
+    </StackPanel>
+
+  </StackPanel>
+
+</Window>