Keep host key trust in the vault, and make it withdrawable

A fingerprint approved once is now approved on every machine and survives a
restart, because host key trust is a vault item type rather than a dictionary
that dies with the process. InMemoryKnownHostStore was what shipped, so the user
was asked to verify a fingerprint on every single connection — which is the gap
most likely to train somebody to click through the one warning that actually
matters. A warning that appears when nothing is wrong teaches that nothing is
ever wrong.

The fourth item type, and like the third it cost no sync logic: a row, an EF
configuration, a migration, a server kind; a secret, a codec, a merge, a cipher,
a repository facade and a session property. One row in the client registry. The
reconciler, the mirror, the repository, the outbox and the pull filter were not
touched. SyncEntityType.KnownHostKey and AadResourceType.KnownHostKey were
already reserved, so neither the contract nor docs/crypto.md changed.

One item per (host, port, algorithm), because a server legitimately offers
several host keys and which one gets negotiated is not ours to predict. Pinning
per endpoint would make an algorithm change indistinguishable from an attack.

The label is derived rather than stored, which is the one place this type
departs from the other three. A user never names a pin — there is nothing to
name it after but the three fields it already has — and a stored label is a
second copy of data that can disagree with the first after a merge. Relabel
returns the secret unchanged, and says why.

The store answers the handshake without touching the disk. SshNetConnectionFactory
calls FindAsync from inside SSH.NET's synchronous HostKeyReceived event, over
.GetAwaiter().GetResult(), which cannot be avoided; doing SQLite I/O plus an AEAD
open per lookup there would put the handshake behind the cache. So decryption
happens in OpenAsync and RefreshAsync — on unlock and after each sync pass,
exactly where the host and key lists already reload — and FindAsync is a
dictionary read under a lock with no await inside it.

That snapshot is where the one real bug in this change lived. Install originally
merged the live pins over the freshly loaded snapshot, to protect a TrustAsync
that had landed while the read was in flight. It would also have resurrected
every pin the user had just forgotten, and stopped a withdrawal made on another
machine from ever taking effect — the store would have healed the deletion back
into existence on every refresh. Replacing wholesale and discarding the read
instead is correct because writes are the rare case: every write bumps a
generation counter, and a refresh whose stamp is stale throws itself away rather
than winning. Nothing found this but reading the method again; it is the kind of
mistake that passes every test written before it, because the test that catches
it is the one the bug tells you to write.

Forgetting is new, and persistence is what made it mandatory rather than
convenient. A mismatch is a hard refusal with no way to continue — deliberately,
and that stays — so pinning a key permanently is also a way to make a
legitimately rebuilt server permanently unreachable. Before this change the pin
died at exit and the problem solved itself; now it does not.

ForgetAsync drops every algorithm for an endpoint, and it is reachable from the
host editor rather than from the warning. Putting it on the mismatch banner would
have made it two clicks from "this may be an attack" to "connect anyway", which
is the affordance the hard refusal exists to deny. The banner already promised
the key could be removed in the host's settings; that promise is now true and
points at the button.

Trust recorded on another machine becomes visible at the next sync pass, not
immediately, and that is a decision rather than an oversight. The failure it
produces is a first-contact prompt for a host a colleague approved a minute ago:
answerable, and self-correcting on the next pass. The opposite trade — polling
the vault on the handshake thread to close a one-minute window — buys nothing
and costs the property above. The dangerous direction is not reachable at all: a
pin recorded here enters the snapshot as part of recording it, so a refresh can
never discard a local trust decision.

The server learns nothing, and this is the item type where the temptation was
real. A plaintext host column would let a known-hosts screen sort and page
without decrypting anything, and it would hand the operator the map of every
user's estate — assembled, as these things are, out of facts that are each
individually harmless. A host row concedes an address only when relay is
switched on and the database refuses to store one otherwise (ADR 0004); there is
no equivalent excuse here. The table has no column to put one in, and the EF
configuration says so where somebody adding it would be standing.

Two things about the migration in this commit are worth knowing, because both
came out of getting it wrong.

It was hand-written first, including its .Designer.cs, and that version is not
what is here. Verifying it turned up something that had been quietly assumed:
Migration_AppliedCleanly_WithNoPendingModelChanges does not check the model
snapshot. It asserts that migrations applied and that none are pending, which a
wrong snapshot satisfies perfectly — the snapshot only matters as the diff base
for the *next* migrations add, so an incorrect one passes the whole suite and
corrupts the following migration instead. The real check is to generate a
throwaway migration and confirm its Up and Down come out empty. They did, and
the generated designer was byte-identical to the transcribed one across all 1255
lines, so the hand-written work was in fact correct.

Then dotnet ef migrations remove --no-build deleted the wrong migration. With
--no-build the tool reads the previously compiled assembly rather than the files
on disk, and the probe had just changed which migration was last, so it removed
AddKnownHostKeyItem and reverted the snapshot. That turned out to leave exactly
the right diff base, so the migration here is EF's own output rather than a
transcription — a better outcome than the one that was interrupted, arrived at
by accident. Never pass --no-build to migrations remove.

Mutation tested, all three sabotages detected: dropping the algorithm from
KnownHostIdentity.For, merging instead of replacing in Install, and pointing
KnownHostKeyCipher at PortForward — which is what a cast from the wire enum's 10
would silently produce. Each is caught both by an assertion about the mechanism
and by a behavioural test that never mentions it; the resource-type sabotage is
caught by the table from d10a38d and nothing else, which is what that table is
for.

The end-to-end slice now approves the real sshd's host key through the vault,
pushes it, and reads it back on the second simulated machine — including a check
that the server learned no address, and that the second machine answers null for
an algorithm never offered.

845 tests green. Zero warnings, dotnet format clean.

Three things are deliberately not fixed. A tombstone queued over a create that
was never pushed is refused by the server as Invalid and parked; that is
pre-existing for all four item types, and the fix belongs in
VaultItemRepository.DeleteAsync rather than here. Deleting a host, or changing
its address, orphans its pins — both are correct as trust decisions, since a pin
describes an endpoint and not a bookmark, but nothing surfaces the leftovers.
And there is no interface listing pins at all: trust is created at the connect
prompt and withdrawn in the host editor. A known-hosts list is where the orphans
would become visible, and it wants the vault column rework first, for the same
reason the credential editor does.

README.md

@@ -132,10 +132,16 @@ private key, then edit a host and pick that key from its **key** dropdown. From
 authenticates with it — on every machine, since the choice travels inside the host's encrypted payload —
 and its password box disappears.
 
-Three of M1's known gaps are visible immediately, so they are worth expecting rather than diagnosing:
-password authentication asks for the password every time, because credentials are not a synced entity type
-yet (keys are — passwords are not); host key trust lasts one session, because known hosts do not live in the
-vault yet; and unlock asks for the passphrase on every launch, because no device key is registered.
+The first time you connect to a host you are asked to check its key fingerprint. That decision is stored in
+the vault, so it is asked once per host rather than once per launch and it reaches your other machines with
+the next sync. If a server is legitimately rebuilt and offers a new key, the connection is refused outright
+with no way to continue from the warning — edit the host and choose **Forget host key**, which is deliberately
+somewhere you have to go on purpose.
+
+Two of M1's known gaps are visible immediately, so they are worth expecting rather than diagnosing: password
+authentication asks for the password every time, because nothing in the interface can create a vault
+credential yet (they do sync — there is just no editor for one); and unlock asks for the passphrase on every
+launch, because no device key is registered.
 
 ### End-to-end verification
 
@@ -148,9 +154,9 @@ dotnet test tests/DodoSSH.SystemTests
 
 It brings up PostgreSQL, Keycloak and an OpenSSH server in containers, applies the committed migrations,
 starts the API as a child process out of its own build output, and then drives the real client: sign in
-through Keycloak, enroll, unlock, create a host, sync it, read it back on a second simulated machine,
-unlock again with no network, and open a shell on the `sshd`. Roughly 25 seconds once the images are
-pulled.
+through Keycloak, enroll, unlock, create an SSH key and a host bound to it, sync them, open a shell on the
+`sshd` and approve its host key at the real first-contact refusal, then read all three back on a second
+simulated machine and unlock again with no network. Roughly 25 seconds once the images are pulled.
 
 What makes it worth its weight is that it consumes the artefacts that ship — the realm file from
 `deploy/keycloak`, the EF migrations, the API's own `appsettings` — rather than a fixture written to match
@@ -198,14 +204,19 @@ off-Windows.
   they are unverified.
   *Verified end to end:* `tests/DodoSSH.SystemTests` drives the whole slice against a real Keycloak, a
   real API, a real PostgreSQL and a real `sshd` — sign-in, the identity-provider key binding, enrollment,
-  offline unlock, a host through the vault to a second machine, and an interactive shell. See
+  offline unlock, a host and an SSH key through the vault to a second machine, an interactive shell, and the
+  host key approved at that shell's prompt reaching the second machine as well. See
   [End-to-end verification](#end-to-end-verification).
 
-  Known gaps in the client, stated rather than implied by the interface: credentials are not a synced
-  entity type yet, so password authentication still asks for the password each time — SSH keys *are*
-  synced, and binding one to a host is the way to connect without typing anything; known host keys live in
-  memory for one session instead of in the vault; and no device key is registered, so the passphrase is
-  needed on every launch until the OS keystore is wired.
+  Known gaps in the client, stated rather than implied by the interface: nothing in the interface can create
+  a vault credential yet, so password authentication still asks for the password each time — SSH keys *are*
+  editable, and binding one to a host is the way to connect without typing anything; and no device key is
+  registered, so the passphrase is needed on every launch until the OS keystore is wired.
+
+  Host key trust *is* in the vault, which is what makes trust-on-first-use worth having: a fingerprint
+  approved on one machine is approved on all of them and survives a restart, and the server cannot drop a
+  pin to force a fresh first-use decision without the item visibly going missing. A changed host key stays a
+  hard refusal with no way past it; withdrawing a pin is a separate, deliberate act in the host's editor.
 
   Binding a key introduced the first payload schema version bump, and it is worth knowing how it behaves:
   a host is written at the *lowest* schema version that can represent it, so only hosts that actually bind