chore(followups): reviewer-recommended tests, comments, and hardening from the remediation reviews

The remediation reviews approved every task but left a tail of small notes.
This lands the gateway-side half of them.

Hardening (behavior changes, all narrow):

- BuildFilteredWriteBulkCommand's unreachable default case failed OPEN: a
  fifth bulk-write kind added upstream without a filter case here would have
  shipped the DENIED entries to the worker while reporting them denied to the
  caller. It now throws UnreachableException.
- SqliteCanonicalAuditStore.ListRecentAsync no longer throws on a row it
  cannot date. The retention sweep deliberately preserves such rows (SQLite's
  datetime() yields NULL, so the DELETE never matches), which guaranteed the
  dashboard's recent-audit view would meet one eventually and lose the whole
  page to it. The row is now reported at DateTimeOffset.MinValue with every
  other column intact, behind an optional logger.
- The audit drain loop's finally now completes the channel writer alongside
  detaching the drain, so a producer that raced past the attached check takes
  the write-through branch instead of stranding its event in a buffer nobody
  reads until shutdown. TryComplete is idempotent, so StopAsync is unaffected.

Tests:

- MapCommandReply ownership (Assert.Same on the inner reply), mirroring the
  existing MapEvent ownership test.
- Redactor key-id length boundary at exactly 64 and 65 characters, pinning
  which way it fails. Nothing validates key-id length at creation, so
  docs/Diagnostics.md's "which no issued key id does" is now stated as the
  heuristic it is.
- ApiKeyFailureLimiter.Reset with a PartitionResolution whose partition was
  evicted between the Check and the Reset: inert, and clears nobody else's
  block.
- Constraint-cache concurrency stress: the cap is enforced by the inserting
  thread, so overshoot must be transient and proportional to the in-flight
  inserters, and the cache must settle at or under the cap.
- ListRecentAsync against a raw-SQL undateable row.

Comment/doc accuracy:

- EventsHubViewerRegistry.ReleaseConnection records that it relies on
  SignalR's default sequential per-connection dispatch
  (MaximumParallelInvocationsPerClient = 1).
- A PERF(followup) note on Invoke's double session resolve and why removing it
  needs a SessionManager overload.
- SessionEventDistributor: the volatile-field comment named the pump as the
  lock-free reader, but the pump's single capture point is inside _replayLock;
  the genuinely lock-free reader is SubscriberCount. OnSubscriberOverflow's
  "cannot be observed here" now excepts the DisposeAsync abandon path. The
  churn test names its ConcurrentDictionary bucket-order assumption and that a
  violation surfaces as a read timeout, not a silent pass.
- The two "restores the sequential drain's behavior" claims (SessionManager,
  docs/Sessions.md) were wrong: the sequential drain leaked too, because
  KillWorkerAsync's entry ThrowIfCancellationRequested aborted the whole loop
  on the first session for zero kills. Reworded to "fixes a leak the
  sequential drain also had", with the sweep-bound/shutdown-unbound
  ParallelOptions asymmetry explained.
- ISessionManager.ShutdownAsync's token doc: it degrades the drain to a kill
  sweep rather than cancelling it, with the bounded overrun stated.
  SessionShutdownHostedService.StopAsync records that its cancellation-logging
  branch is now unreachable.

docs/Diagnostics.md

@@ -101,8 +101,12 @@ correlate a log entry to a specific principal.
 A scheme word survives only when it is one of the recognized authorization schemes (`Bearer`,
 `Basic`, `Digest`, `Negotiate`, `NTLM`, `ApiKey`, `Token`). An unrecognized leading word is as likely
 to be credential material as it is to be a scheme, so it is dropped along with the rest. The key id is
-also dropped when it runs longer than 64 characters, which no issued key id does — a long run before
-the first `_` is secret material, not an identifier.
+also dropped when it runs longer than 64 characters — a long run before the first `_` is more likely to
+be secret material than an identifier. Neither key-creation path (`ApiKeyAdminCommandLineParser.IsValidKeyId`,
+`DashboardApiKeyManagementService.ValidateKeyId`) enforces a length, so this is a redaction heuristic
+rather than a guarantee: operators should keep key ids under 64 characters, or the id stops appearing
+in logs and only the `mxgw_[redacted]` shape survives. The direction of the failure is deliberate —
+losing an identifier is cheap, logging a secret is not.
 
 The parse is span-based (no regex, no `Split` allocation): the value is split once at the first space,
 and the key id is read up to the first `_` of the remainder.

docs/Sessions.md

@@ -215,7 +215,9 @@ Splitting the phases moves the TOCTOU re-check earlier, and that is an accepted
 
 A close that throws no longer abandons the rest of the selected set: the sweep attempts every selected session, captures the first failure, and rethrows it once the pass is done so `SessionLeaseMonitorHostedService` still logs the sweep failure as before. Because the closes run concurrently, *which* failure surfaces when several fail in one pass is nondeterministic; the log line is the diagnostic, not the identity of the exception.
 
-`ShutdownAsync` drains sessions with the same bounded fan-out and the same per-session catch → `KillWorkerAsync` fallback, with two rules that keep a stop deadline from turning into leaked workers. First, its body is **exception-total** — nothing escapes it — because `Parallel.ForEachAsync` cancels the token handed to the sibling bodies as soon as one body throws, which would abort in-flight graceful shutdowns *and* make their kill fallback fail instantly on the freshly cancelled token. Second, the drain loop is deliberately **not** bound to the caller's `CancellationToken` and the kill fallback runs on `CancellationToken.None`: a cancelled `ParallelOptions` token stops dispatching the remaining sessions entirely, whereas the sequential drain this replaced let every remaining session fail its graceful close fast and still kill its worker. The token is passed to the graceful close instead, so a host stop deadline turns the drain into a kill sweep rather than into a leak. This matters because nothing reattaches to a leaked worker — a restarted gateway terminates orphans (see [Design Decisions](DesignDecisions.md)).
+`ShutdownAsync` drains sessions with the same bounded fan-out and the same per-session catch → `KillWorkerAsync` fallback, with two rules that keep a stop deadline from turning into leaked workers. First, its body is **exception-total** — nothing escapes it — because `Parallel.ForEachAsync` cancels the token handed to the sibling bodies as soon as one body throws, which would abort in-flight graceful shutdowns *and* make their kill fallback fail instantly on the freshly cancelled token. Second, the drain loop is deliberately **not** bound to the caller's `CancellationToken` and the kill fallback runs on `CancellationToken.None`: a cancelled `ParallelOptions` token stops dispatching the remaining sessions entirely, so the untried tail would be neither closed nor killed. This **fixes a leak the sequential drain also had**, rather than restoring the sequential drain's behavior — there the kill fallback ran on the caller's already-cancelled token, and `KillWorkerAsync`'s entry `ThrowIfCancellationRequested` threw out of the loop on the very first session, producing zero kills. The token is passed to the graceful close only, so a host stop deadline turns the drain into a kill sweep rather than into a leak. This matters because nothing reattaches to a leaked worker — a restarted gateway terminates orphans (see [Design Decisions](DesignDecisions.md)).
+
+The asymmetry with `CloseExpiredLeasesAsync` — whose `ParallelOptions` *is* token-bound — is intentional: the sweep is periodic maintenance, so a pass abandoned on cancellation loses nothing permanently (the next pass re-selects, and `ShutdownAsync` backstops it), whereas the shutdown drain is terminal and must not be abandoned partway.
 
 **Stranded-`Closing` bound.** A sweep pass that is cancelled after selection leaves its unclosed selections in `Closing` with close already started. `IsFaultedReapableCore` requires `state == Faulted`, so a session selected under `FaultedReason` and stranded this way is not re-selected as faulted; it is swept only when its normal lease expires (up to `MxGateway:Sessions:DefaultLeaseSeconds`, default 1800 s), since `IsLeaseExpiredCore` and `IsDetachGraceExpiredCore` are state-agnostic. This bound is documented rather than closed with a re-selection clause: the sweep's only caller cancels on the host's `stoppingToken`, so the very next thing that runs is `ShutdownAsync`, which drains (or kills) the whole registry — and any worker that still survives that is terminated as an orphan on the next gateway start. Adding a "`Closing` and close-started" re-selection clause would also have to distinguish an abandoned close from one that is merely still in flight, which would weaken the single invariant that makes the parallel close phase safe.
 

docs/plans/2026-08-15-perf-review-remediation.md

@@ -617,6 +617,7 @@ Commit anything found: `docs: remediation plan doc sweep`
 | Control-frame completion coupled to event batch drain | Documented, bounded (≤128 frames) behavior of the two-class writer design; revisit only if heartbeat latency shows up in metrics. |
 | Blazor pages' loopback SignalR hop | Works correctly; in-process `WatchSnapshotsAsync` consumption is a dashboard refactor with payoff only at viewer counts the product doesn't target. |
 | Event-path triple async-iterator flattening | LOW-rated; touches the most invariant-dense code in the gateway for two `MoveNextAsync` hops per event. Reconsider after Tasks 3/4 land and if profiling still shows it. |
+| `SessionEventDistributor._subscribers` `ConcurrentDictionary` → plain `Dictionary` (Task 25 / Task 3 review) | Every mutation is already inside `_lifecycleLock`, so the concurrent type buys nothing. Behavior-neutral refactor with no measurable win, proposed after the windev gate was already green — not worth re-running the verification matrix for. Comment cleanups from the same review landed; this swap did not. |
 
 ## Execution notes for the orchestrator