Resolve Tests-003, -004, -005, -006 code-review findings

Tests-003: temp auth-DB directories leaked under %TEMP%. Added the
TempDatabaseDirectory IDisposable helper (clears the Sqlite connection pool,
then recursively deletes); SqliteAuthStoreTests and ApiKeyAdminCliRunnerTests
now dispose every directory they create.

Tests-004: added end-to-end coverage composing the real authorization
interceptor in front of the real MxAccessGatewayService, plus scope-resolver
tests confirming an unmapped request type fails closed to the admin scope.

Tests-005: added coverage for a worker faulting mid-command — a pipe
disconnect and a worker fault while an InvokeAsync is in flight both fail the
pending invoke. No product change needed.

Tests-006 (re-triaged): the flaky ReadLoop_WhenClientFaults_KillsOwnedWorkerProcess
is a test race, not a product bug — the kill runs synchronously inside
SetFaulted. Rewrote it to await FakeWorkerProcess exit deterministically, and
replaced fixed Task.Delay timing in the late-reply and heartbeat tests with
FIFO ordering and an injected ManualTimeProvider.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

code-reviews/Tests/findings.md

@@ -7,7 +7,7 @@
 | Review date | 2026-05-18 |
 | Commit reviewed | `6c64030` |
 | Status | Reviewed |
-| Open findings | 10 |
+| Open findings | 6 |
 
 ## Checklist coverage
 
@@ -65,13 +65,13 @@
 | Severity | Medium |
 | Category | Performance & resource management |
 | Location | `src/MxGateway.Tests/Security/Authentication/SqliteAuthStoreTests.cs:170-176`, `src/MxGateway.Tests/Security/Authentication/ApiKeyAdminCliRunnerTests.cs:252-258` |
-| Status | Open |
+| Status | Resolved |
 
 **Description:** `CreateTempDatabasePath` creates a fresh directory under `%TEMP%\mxgateway-auth-tests\<guid>` (and `...-cli-tests`) for every test but nothing ever deletes it. `WorkerProcessLauncherTests.TestDirectory` correctly implements `IDisposable` and cleans up; these two do not. SQLite connection pooling can also keep the `.db` handle open after the test. Over many CI runs this leaks temp files and open handles.
 
 **Recommendation:** Wrap the temp directory in an `IDisposable`/`IAsyncDisposable` helper (as `WorkerProcessLauncherTests` does) and call `SqliteConnection.ClearAllPools()` before deletion, or use `Microsoft.Data.Sqlite` in-memory mode where a real file is not needed.
 
-**Resolution:** _(open)_
+**Resolution:** Resolved 2026-05-18: confirmed root cause — both `CreateTempDatabasePath` helpers created `%TEMP%` directories with no cleanup, and `Microsoft.Data.Sqlite` pools connections by default so the `.db` handle outlives the test. Added a shared `TempDatabaseDirectory` (`src/MxGateway.Tests/Security/Authentication/TempDatabaseDirectory.cs`) `IDisposable` helper that calls `SqliteConnection.ClearAllPools()` and recursively deletes its directory. `SqliteAuthStoreTests` and `ApiKeyAdminCliRunnerTests` now implement `IDisposable`, track every directory created via `CreateTempDatabasePath`, and dispose them after each test. All affected tests still pass.
 
 ### Tests-004
 
@@ -80,13 +80,13 @@
 | Severity | Medium |
 | Category | Testing coverage |
 | Location | `src/MxGateway.Tests/Security/Authorization/GatewayGrpcAuthorizationInterceptorTests.cs` |
-| Status | Open |
+| Status | Resolved |
 
 **Description:** The authorization interceptor and `MxAccessGatewayService` are each tested in isolation, but no test composes the interceptor in front of the real service to confirm scope enforcement gates real RPCs end-to-end. A wiring mistake — interceptor not registered, or a new RPC added without a scope mapping in `GatewayGrpcScopeResolver` — would pass every existing test. `GatewayGrpcScopeResolverTests` also only checks an enumerated allow-list; it never asserts an unmapped request type fails closed.
 
 **Recommendation:** Add an end-to-end test that runs `OpenSession`/`Invoke` through the interceptor+service composition with insufficient scope and asserts `PermissionDenied`; add a `GatewayGrpcScopeResolver` test asserting an unknown/unmapped request type throws or denies rather than returning a permissive default.
 
-**Resolution:** _(open)_
+**Resolution:** Resolved 2026-05-18: confirmed the coverage gap. Added three interceptor+service composition tests to `GatewayGrpcAuthorizationInterceptorTests` that run the real `GatewayGrpcAuthorizationInterceptor` continuation into a real `MxAccessGatewayService`: `InterceptorComposedWithService_OpenSessionMissingScope_DeniesBeforeServiceRuns` (asserts `PermissionDenied` and `OpenSessionCount == 0`), `InterceptorComposedWithService_OpenSessionWithScope_RunsServiceWithIdentity` (service runs and observes the interceptor-pushed identity), and `InterceptorComposedWithService_InvokeWriteCommandWithReadScope_DeniesBeforeServiceRuns` (a `Write` command with only `invoke:read` is denied). Added two `GatewayGrpcScopeResolverTests`: `ResolveRequiredScope_UnmappedRequestType_FailsClosedToAdminScope` confirms an unmapped request type resolves to the most-restrictive `Admin` scope (the resolver's `_ => GatewayScopes.Admin` default already fails closed — no product bug), and `ResolveRequiredScope_UnknownInvokeCommandKind_ReturnsInvokeReadScope` confirms an unknown command kind does not silently grant write/admin access.
 
 ### Tests-005
 
@@ -95,13 +95,13 @@
 | Severity | Medium |
 | Category | Testing coverage |
 | Location | `src/MxGateway.Tests/Gateway/Grpc/EventStreamServiceTests.cs:239-261`, `src/MxGateway.Tests/Gateway/Sessions/SessionManagerTests.cs` |
-| Status | Open |
+| Status | Resolved |
 
 **Description:** Worker-crash handling is only tested as a clean terminal exception from `ReadEventsAsync` or a pre-set `ShutdownException`. There is no test for a worker that faults mid-command — an `InvokeAsync` in flight when the pipe/worker dies — which is a core fault-handling path of the two-process design. `WorkerClientTests` covers pipe-disconnect faulting the read loop, but not the interaction where a pending `InvokeAsync` task observes the fault and surfaces a meaningful error code.
 
 **Recommendation:** Add a `WorkerClient`/`SessionManager` test that disposes the worker pipe (or emits a `WorkerFault`) while an `InvokeAsync` is pending, and assert the invoke task fails with a `WorkerClientException`/`SessionManagerException` carrying the worker-faulted error code.
 
-**Resolution:** _(open)_
+**Resolution:** Resolved 2026-05-18: confirmed the coverage gap and confirmed the product path already handles it correctly (`WorkerClient.ReadLoopAsync` → `SetFaulted` → `CompletePendingCommands(fault)` fails every pending command with the fault exception). Added two `WorkerClientTests`: `InvokeAsync_WhenPipeDisconnectsMidCommand_FailsPendingInvokeWithPipeDisconnected` (worker reads the command then disposes its pipe side; the pending invoke task fails with `WorkerClientErrorCode.PipeDisconnected`) and `InvokeAsync_WhenWorkerFaultsMidCommand_FailsPendingInvokeWithWorkerFaulted` (worker emits a `WorkerFault` envelope while the invoke is pending; the task fails with `WorkerClientErrorCode.WorkerFaulted`). Both also assert the client transitions to `Faulted`. No product change needed.
 
 ### Tests-006
 
@@ -110,13 +110,15 @@
 | Severity | Medium |
 | Category | Concurrency & thread safety |
 | Location | `src/MxGateway.Tests/Gateway/Workers/WorkerClientTests.cs:76`, `src/MxGateway.Tests/Gateway/Workers/FakeWorkerHarnessTests.cs:122` |
-| Status | Open |
+| Status | Resolved |
 
 **Description:** Several tests rely on fixed `Task.Delay` values: `WorkerClientTests.InvokeAsync_WithLateReply…` waits a hard-coded 50 ms after writing a late reply before issuing the second command, and the heartbeat tests use a 20 ms delay to make timestamps strictly increase. On a slow CI agent the 50 ms delay can be insufficient, and `DateTimeOffset.UtcNow` resolution can make the 20 ms heartbeat-advance assertion flaky.
 
 **Recommendation:** Replace fixed delays with the existing `WaitUntilAsync` condition polling, and inject a controllable `TimeProvider` for heartbeat-timestamp comparisons instead of relying on wall-clock advance.
 
-**Resolution:** _(open)_
+**Re-triage note:** The brief flagged `ReadLoop_WhenClientFaults_KillsOwnedWorkerProcess` as "a real `WorkerClient` fault→kill bug". On inspection it is **not a product bug** — it is a test race. `WorkerClient.SetFaulted` publishes the `Faulted` state under lock *before* calling `KillOwnedProcess`, so the old test's `WaitUntilAsync(() => client.State == Faulted)` could return between those two statements and observe `process.KillCount == 0`. The kill itself always runs synchronously inside `SetFaulted`, and `ShutdownAsync`/`DisposeAsync` re-issue an idempotent kill, so no real consumer relies on "state==Faulted implies process dead". The fix is therefore a test-quality fix (correctly Medium / Concurrency), not a product fix.
+
+**Resolution:** Resolved 2026-05-18: (1) Made `ReadLoop_WhenClientFaults_KillsOwnedWorkerProcess` deterministic — it now `await`s `FakeWorkerProcess.WaitForExitAsync` (the `TaskCompletionSource` completed inside `Kill()`), which completes exactly when the kill runs, eliminating the state-polling race; verified by running it five times in isolation (5/5 pass). (2) Removed the fixed 50 ms `Task.Delay` from `InvokeAsync_WithLateReply_IgnoresLateReplyAndKeepsClientReady` — the stale reply and the second reply are now sent in pipe (FIFO) order, so the read loop discards the stale reply before the second reply with no timing window. (3) Replaced the 20 ms `Task.Delay` heartbeat-advance hacks in `WorkerClientTests.ReadLoop_WhenHeartbeatArrives_UpdatesLastHeartbeatAndWorkerProcess` and `FakeWorkerHarnessTests.SendHeartbeatAsync_UpdatesClientHeartbeatState` with an injected `ManualTimeProvider` advanced by a fixed `TimeSpan`; both tests now assert the exact post-advance timestamp instead of `>` against wall-clock drift.
 
 ### Tests-007