mxaccessgw/code-reviews/Worker/findings.md

# Code Review — Worker

| Field | Value |
|---|---|
| Module | `src/MxGateway.Worker` |
| Reviewer | Claude Code |
| Review date | 2026-05-18 |
| Commit reviewed | `6c64030` |
| Status | Reviewed |
| Open findings | 0 |

## Checklist coverage

| # | Category | Result |
|---|---|---|
| 1 | Correctness & logic bugs | Issues found: heartbeat loop sleeps before first beat (Worker-002), `ProcessCommandAsync` state race drops replies (Worker-003), watchdog/heartbeat state inconsistency (Worker-004), double-dispose path (Worker-006), plus Worker-010/011/015. |
| 2 | mxaccessgw conventions | Issue found: Worker-007 (reflection-based COM invocation bypasses the typed interface contract). |
| 3 | Concurrency & thread safety | Issues found: Worker-001 (`WnWrapAlarmConsumer` timer fires COM off the STA), Worker-008 (consumer factory STA-affinity not enforced). |
| 4 | Error handling & resilience | Issue found: Worker-005 (`OnPoll` silently swallows all poll failures). |
| 5 | Security | No secret logging (redaction applied); inbound frame validation reasonable. No issues found. |
| 6 | Performance & resource management | Issue found: Worker-009 (per-frame `byte[]` allocations on the hot event path). COM release is correct. |
| 7 | Design-document adherence | Code matches `WorkerSta.md`/`WorkerFrameProtocol.md`; stale alarm-path docs (Worker-012). |
| 8 | Code organization & conventions | Issue found: Worker-014 (`AlarmCommandHandler.cs` declares two public types in one file). |
| 9 | Testing coverage | Issue found: Worker-013 (`StaMessagePump` has no direct tests; poll-loop lifecycle untested). |
| 10 | Documentation & comments | Issue found: Worker-012 (stale "future PR / A.3" comments now describe shipped code). |

## Findings

### Worker-001

| Field | Value |
|---|---|
| Severity | High |
| Category | Concurrency & thread safety |
| Location | `src/MxGateway.Worker/MxAccess/WnWrapAlarmConsumer.cs:204-207` |
| Status | Resolved |

**Description:** When constructed with `pollIntervalMilliseconds > 0`, `Subscribe` starts a `System.Threading.Timer` whose `OnPoll` callback runs `PollOnce()` — which calls `wwAlarmConsumerClass.GetXmlCurrentAlarms2` — on a thread-pool thread. The wnwrap CLSID is registered `ThreadingModel=Apartment`; calling its methods off the owning STA violates the hard rule that all COM calls happen on the dedicated STA thread, and can deadlock on cross-apartment marshaling when the STA is not pumping. The production path (default constructor, interval 0) is safe, but the public 3-arg constructor leaves this footgun callable, and tests/live-smoke use it.

**Recommendation:** Remove the internal `Timer` entirely (production already drives `PollOnce` from the STA), or document and gate it so it can only be used from an STA thread. At minimum, make the timer-driven mode unreachable from any production wiring.

**Resolution:** 2026-05-18 — Removed the off-STA timer infrastructure from `WnWrapAlarmConsumer`: the `Timer? pollTimer` and `pollIntervalMs` fields, the `DefaultPollIntervalMilliseconds` constant, the `OnPoll` callback, the timer-arming arm in `Subscribe`, and the timer disposal block in `Dispose`. The `pollIntervalMilliseconds` parameter is gone from both public constructors (the test-seam ctor is now 2-arg: `wwAlarmConsumerClass` + `maxAlarmsPerFetch`), so the off-STA footgun is structurally unreachable. `PollOnce()` remains the public STA-driven entry point. The stale "poll … on a timer below" comment was corrected. Verified by the regression tests `WnWrapAlarmConsumer_has_no_internal_timer_field` and `WnWrapAlarmConsumer_exposes_no_poll_interval_constructor_parameter`; the `AlarmsLiveSmokeTests` call site was updated to the 2-arg constructor.

### Worker-002

| Field | Value |
|---|---|
| Severity | High |
| Category | Correctness & logic bugs |
| Location | `src/MxGateway.Worker/Ipc/WorkerPipeSession.cs:545-549` |
| Status | Resolved |

**Description:** `RunHeartbeatLoopAsync` calls `await Task.Delay(_sessionOptions.HeartbeatInterval, ...)` before sending the first heartbeat. The gateway therefore receives no heartbeat for the first full interval (default 5s) after the worker reaches `Ready`. If the gateway's liveness watchdog expects a heartbeat sooner, a healthy worker can be misclassified as hung at startup.

**Recommendation:** Send an initial heartbeat immediately on entering the loop, or move the `Task.Delay` to the end of the loop body.

**Resolution:** 2026-05-18 — Restructured `RunHeartbeatLoopAsync` so the `Task.Delay(HeartbeatInterval)` is applied between beats only, not before the first. A `firstBeat` guard skips the delay on the initial iteration, so the gateway sees a heartbeat as soon as the worker is `Ready`; cancellation behavior is preserved (the loop still observes the token and the delay still throws on cancellation). Verified by the regression test `RunAsync_SendsFirstHeartbeatImmediatelyOnEnteringLoop`. Three pre-existing tests (`WorkerPipeClientTests.RunAsync_ConnectsToPipeAndCompletesHandshake`, `WorkerPipeClientTests.RunAsync_RetriesUntilPipeServerAppears`, `WorkerPipeSessionTests.RunAsync_WhenCommandThrowsAfterShutdown_DropsLateFaultAndWritesShutdownAck`) assumed strict frame ordering and were updated to skip the now-interleaved first heartbeat while still asserting the same shutdown-ack behavior.

### Worker-003

| Field | Value |
|---|---|
| Severity | High |
| Category | Correctness & logic bugs |
| Location | `src/MxGateway.Worker/Ipc/WorkerPipeSession.cs:399-403`, `:416-419` |
| Status | Resolved |

**Description:** `ProcessCommandAsync` checks `_state` after `DispatchAsync` completes and silently `return`s without writing a `WorkerCommandReply` (or fault) when `_state` is not `Ready`/`ExecutingCommand`. `_state` is a plain field mutated from multiple tasks (heartbeat loop, event-drain loop, shutdown). A command that completes successfully while `_state` has transitioned will have its reply dropped with no diagnostic, and the gateway's correlation-id wait then hangs until its own timeout. The `_state` read is also not synchronized.

**Recommendation:** Always attempt to write the reply/fault for an in-flight command, or explicitly reject in-flight commands with a `Canceled`/`WorkerUnavailable` reply during state transitions. Make `_state` access thread-safe (volatile or locked).

**Resolution:** 2026-05-18 — Both silent-drop `return` sites in `ProcessCommandAsync` (the post-`DispatchAsync` success path and the exception path) now call a new `LogCommandResultDropped` helper before returning. The helper logs an Information event named `WorkerCommandResultDropped` via the session's `IWorkerLogger`, carrying the command's `correlation_id` plus `command_method` and `worker_state`, so a stuck gateway correlation-id wait is now traceable. The `_state` field was made `volatile` (`WorkerState` is an int-backed protobuf enum, so volatile is valid) so cross-thread reads observe the latest value without tearing; this is a low-risk, non-behavioral change and did not destabilize any test. Verified by the regression test `RunAsync_WhenReplyIsDroppedAfterShutdown_LogsDiagnostic`.

### Worker-004

| Field | Value |
|---|---|
| Severity | Medium |
| Category | Correctness & logic bugs |
| Location | `src/MxGateway.Worker/Ipc/WorkerPipeSession.cs:565-588` |
| Status | Resolved |

**Description:** After `ReportWatchdogFaultIfNeededAsync` sends an `StaHung` fault, the heartbeat loop continues sending normal heartbeats with `State` derived from `_state`, which the watchdog path never sets to `Faulted`. The heartbeat then keeps reporting a non-faulted state that contradicts the fault just sent.

**Recommendation:** Set `_state = WorkerState.Faulted` (thread-safely) when the watchdog fault fires so heartbeat state and fault stay consistent.

**Resolution:** 2026-05-18 — `ReportWatchdogFaultIfNeededAsync` now sets `_state = WorkerState.Faulted` immediately after `_watchdogFaultSent = true` and before the `StaHung` fault is written, so the next heartbeat reports `Faulted` instead of contradicting the fault. `_state` is already `volatile` (Worker-003), so the cross-thread write from the heartbeat loop is observed correctly by the heartbeat's own `CreateHeartbeat` read; no further locking is required. Verified by the regression test `WorkerPipeSessionTests.RunAsync_AfterWatchdogFault_HeartbeatReportsFaultedState`, which uses a stale-activity snapshot with an empty current-command correlation id so the heartbeat `State` is derived from `_state` rather than forced to `ExecutingCommand`.

### Worker-005

| Field | Value |
|---|---|
| Severity | Medium |
| Category | Error handling & resilience |
| Location | `src/MxGateway.Worker/MxAccess/MxAccessStaSession.cs:205-258` (production alarm poll loop) |
| Status | Resolved |

**Description:** `OnPoll` catches every exception from `PollOnce()` and discards it (`_ = ex;`). The production poll path (`MxAccessStaSession.RunAlarmPollLoopAsync` → `AlarmCommandHandler.PollOnce` → `AlarmDispatcher.PollOnce` → `consumer.PollOnce()`) has no fault recording either. A permanently failing alarm provider (e.g. `GetXmlCurrentAlarms2` returning `E_FAIL`, malformed XML throwing in `XmlDocument.LoadXml`) is therefore completely silent — no fault on the event queue, no log.

**Recommendation:** Route poll failures to `MxAccessEventQueue.RecordFault` (or a logger) so a broken alarm subscription becomes observable. Update the now-stale comment.

**Re-triage:** The cited location `WnWrapAlarmConsumer.cs:297-313` and the `OnPoll` callback no longer exist as of this branch — Worker-001 removed the off-STA `Timer` and its `OnPoll` callback entirely. The substantive concern still held, however: the **production** poll path in `MxAccessStaSession.RunAlarmPollLoopAsync` caught only `OperationCanceledException`, `ObjectDisposedException`, and `InvalidOperationException`. A genuine poll failure (`COMException` from `GetXmlCurrentAlarms2`, a malformed-XML `XmlException`) escaped uncaught, faulted the never-awaited `Task.Run` poll task, and was silently lost — exactly the silent-failure the finding describes. The finding was re-pointed at the live location and fixed there rather than at the removed `OnPoll`.

**Resolution:** 2026-05-18 — `RunAlarmPollLoopAsync` gained a trailing `catch (Exception exception)` arm after the three graceful-stop catches. A real alarm-poll failure is now converted to a `WorkerFault` (category `MxaccessEventConversionFailed`, carrying the exception type and, for a `COMException`, its `HResult`) by the new `CreateAlarmPollFault` helper and recorded on the session's `MxAccessEventQueue` via `RecordFault`. The worker's event-drain loop drains that fault and forwards it to the gateway, so a broken alarm subscription is now observable on the IPC fault path instead of vanishing. The poll loop still stops after the failure (the subscription is dead). No new proto enum value was added — `MxaccessEventConversionFailed` is the closest existing alarm-path category, avoiding a contracts regeneration across all clients. Verified by the regression test `MxAccessStaSessionTests.RunAlarmPollLoop_WhenPollOnceThrows_RecordsFaultOnEventQueue`.

### Worker-006

| Field | Value |
|---|---|
| Severity | Medium |
| Category | Correctness & logic bugs |
| Location | `src/MxGateway.Worker/Ipc/WorkerPipeSession.cs:117-124`, `src/MxGateway.Worker/MxAccess/MxAccessStaSession.cs:386-491` |
| Status | Resolved |

**Description:** `RunAsync`'s `finally` calls `_runtimeSession?.Dispose()` unless `_shutdownTimedOut`. On the normal path `ShutdownGracefullyAsync` already disposed the STA runtime, so re-entering `Dispose()` is a harmless no-op only because `ShutdownGracefullyAsync` reached its end and set `disposed = true`. If `ShutdownGracefullyAsync` throws `TimeoutException` after partial teardown with `_shutdownTimedOut` set, the session is never disposed at all — the `finally` skips it — leaking the STA thread and COM object, leaving cleanup to rely solely on process exit.

**Recommendation:** Make the dispose decision explicit and confirm process exit always follows a timed-out shutdown; otherwise dispose defensively. At minimum document why disposal is deliberately skipped on timeout.

**Resolution:** 2026-05-18 — `RunAsync`'s `finally` now always calls `_runtimeSession?.Dispose()`; the `if (!_shutdownTimedOut)` guard and the `_shutdownTimedOut` field (which had become write-only) were removed. `MxAccessStaSession.Dispose` is idempotent (`if (disposed) return`) and bounded — each STA join is capped with `Wait(TimeSpan.FromSeconds(2))` — so re-entering it on the normal path (where `ShutdownGracefullyAsync` already disposed the runtime) is a harmless no-op, while on the timed-out path it is now the only thing that reclaims the STA thread and releases the MXAccess COM object. The previous behaviour leaked both on a shutdown timeout and relied solely on process exit. A code comment in the `finally` block documents the reasoning. Verified by the regression test `WorkerPipeSessionTests.RunAsync_WhenShutdownTimesOut_StillDisposesRuntimeSession`, which forces a `TimeoutException` from `ShutdownGracefullyAsync` and asserts the runtime session is disposed before `RunAsync` rethrows.

### Worker-007

| Field | Value |
|---|---|
| Severity | Medium |
| Category | mxaccessgw conventions |
| Location | `src/MxGateway.Worker/MxAccess/MxAccessComServer.cs:130-150` |
| Status | Resolved |

**Description:** `Invoke` uses late-bound `Type.InvokeMember` reflection as a fallback when the COM object does not cast to `ILMXProxyServer*`. In production the object is always `LMXProxyServerClass`, so the reflection path exists only for test doubles — it is dead/untested code on the production path and obscures the interface contract. `params object[] arguments` also boxes value-type handles on every call.

**Recommendation:** Drop the reflection fallback and require the COM object to implement the interface (tests can supply a typed fake), or clearly mark the fallback as test-only.

**Re-triage:** The finding's claim that the reflection path is "dead/untested code" is partly inaccurate — it was in fact the path exercised by the entire `MxAccessCommandExecutorTests` suite, whose `FakeMxAccessComObject` did not implement any typed interface. So the reflection fallback was test-only but *not* untested. The convention concern (bypassing the typed interface contract, boxing value-type handles) is valid, so the fix follows the recommendation's first option.

**Resolution:** 2026-05-18 — The late-bound `Type.InvokeMember` reflection fallback and its `params object[]`-boxing `Invoke` helper were removed from `MxAccessComServer`. Each adapter method now takes one of two typed paths: an `is IMxAccessServer` fast path (test fakes implement `IMxAccessServer` directly) and the production path that casts to the typed `ILMXProxyServer` / `ILMXProxyServer3` / `ILMXProxyServer4` COM interfaces via new `AsProxyServer*` helpers. A COM object implementing neither now fails fast with a clear `InvalidOperationException` naming the missing interface, instead of an opaque late-bound call. The test seam was migrated accordingly: `MxAccessCommandExecutorTests.FakeMxAccessComObject` now declares `: IMxAccessServer` (its method signatures already matched the interface exactly, so no behavioural change). Verified by the new `MxAccessComServerTests` (typed-server routing, untyped-object rejection, original-exception propagation — no more `TargetInvocationException` wrapping) plus the unchanged, still-passing `MxAccessCommandExecutorTests` suite which now exercises the typed `IMxAccessServer` path.

### Worker-008

| Field | Value |
|---|---|
| Severity | Medium |
| Category | Concurrency & thread safety |
| Location | `src/MxGateway.Worker/MxAccess/MxAccessStaSession.cs:205-249`, `:429-447` |
| Status | Resolved |

**Description:** `RunAlarmPollLoopAsync` correctly marshals `handler.PollOnce()` onto the STA via `staRuntime.InvokeAsync`, and the cancel/await/dispose ordering in `ShutdownGracefullyAsync` is sound. However, nothing enforces that the `consumerFactory` and all `IMxAccessAlarmConsumer` calls run on the STA thread; a future caller could break STA affinity silently.

**Recommendation:** Add an assertion or documented invariant that the consumer factory and all `IMxAccessAlarmConsumer` calls run on the STA thread, mirroring the existing `MxAccessSession.CreationThreadId` pattern.

**Resolution:** 2026-05-18 — `MxAccessStaSession` now records the STA thread id (`alarmConsumerThreadId`) at the point the alarm-command-handler factory is invoked — which already runs inside `staRuntime.InvokeAsync` during `StartAsync`, mirroring the `MxAccessSession.CreationThreadId` capture. `RunAlarmPollLoopAsync`'s marshalled poll lambda now calls `EnsureOnAlarmConsumerThread()` before `handler.PollOnce()`, asserting the poll runs on the recorded STA thread. The check is delegated to a new `internal static` guard `AssertOnAlarmConsumerThread(int? expected, int actual)` that throws a descriptive `InvalidOperationException` on an affinity violation and is a no-op when the consumer thread is unrecorded (no alarm handler configured). Making the guard `static` and `internal` keeps it directly unit-testable. The STA-affinity invariant is documented in the guard's XML doc. Verified by the regression tests `MxAccessStaSessionTests.AssertOnAlarmConsumerThread_WhenOffOwningThread_Throws` and `AssertOnAlarmConsumerThread_OnOwningThreadOrUnset_DoesNotThrow`.

### Worker-009

| Field | Value |
|---|---|
| Severity | Low |
| Category | Performance & resource management |
| Location | `src/MxGateway.Worker/Ipc/WorkerFrameReader.cs:31,49`, `src/MxGateway.Worker/Ipc/WorkerFrameWriter.cs:57-58` |
| Status | Resolved |

**Description:** Every frame read allocates a fresh 4-byte length buffer and a payload `byte[]`; every write allocates `ToByteArray()` plus a 4-byte prefix. On the hot event-drain path (batches of up to 128 `WorkerEvent` frames every 25 ms) this produces steady gen-0 garbage. `WorkerFrameWriter` also effectively serializes twice (`CalculateSize()` then `ToByteArray()`).

**Recommendation:** Reuse a pooled buffer / `ArrayPool<byte>` for the length prefix and payload, and write directly into a pooled buffer using `CodedOutputStream`. Low priority unless event throughput is high.

**Resolution:** 2026-05-18 — `WorkerFrameWriter.WriteAsync` now serializes the envelope exactly once into a single frame buffer that carries the 4-byte length prefix followed by the payload, via `envelope.WriteTo(new Span<byte>(frame, sizeof(uint), payloadLength))`. This eliminates the redundant second serialization pass (`ToByteArray()` re-runs `CalculateSize()` internally), the separate length-prefix array, and the separate prefix `WriteAsync`/extra `FlushAsync` round. `WorkerFrameReader.ReadAsync` now rents its payload buffer from `ArrayPool<byte>.Shared` and returns it in a `finally` once `WorkerEnvelope.Parser.ParseFrom(payload, 0, length)` has copied what it needs; `ReadExactlyOrThrowAsync` gained an explicit `count` parameter so it honours the logical frame length rather than the (possibly larger) rented buffer length. The 4-byte length-prefix buffer is left as a per-call stack-sized allocation — pooling a 4-byte array is not worthwhile. Verified by the new regression test `WorkerFrameProtocolTests.ReadAsync_WithVaryingFrameSizes_ParsesEachFrameExactly`, which reads a large frame followed by a small frame through one reader to prove the pooled buffer is sliced to each frame's own length and never leaks stale trailing bytes; the existing round-trip, malformed-payload, and concurrent-write tests continue to pass.

### Worker-010

| Field | Value |
|---|---|
| Severity | Low |
| Category | Correctness & logic bugs |
| Location | `src/MxGateway.Worker/Conversion/VariantConverter.cs:204-226` |
| Status | Resolved |

**Description:** `ConvertInt64Scalar` is reached for `TypeCode.UInt32` and `TypeCode.Int64`. For a `uint` with `expectedDataType == MxDataType.Time`, the value is treated as a Windows `FILETIME` via `DateTime.FromFileTimeUtc(longValue)`; a 32-bit FILETIME is never a valid full FILETIME, so this silently produces a near-epoch timestamp rather than a raw/diagnostic value. Unlikely in practice but a silent misconversion.

**Recommendation:** Only apply the `MxDataType.Time` FILETIME projection for 64-bit source types; for `uint` fall through to integer or raw.

**Resolution:** 2026-05-18 — `ConvertInt64Scalar`'s `MxDataType.Time` FILETIME projection is now gated on `value is long`. A genuine 64-bit `long` still projects to a `Timestamp` via `DateTime.FromFileTimeUtc`; a 32-bit `uint` — which can only hold the low half of a FILETIME — now falls through to the integer projection (`DataType = Integer`, `Int64Value`) instead of silently producing a bogus near-1601 timestamp. Verified by the regression test `VariantConverterTests.Convert_WithUInt32AndExpectedTime_DoesNotProjectFileTime`; the existing `Convert_WithFileTimeAndExpectedTime_ProjectsTimestamp` (a `long` FILETIME) continues to pass, confirming the 64-bit path is unchanged.

### Worker-011

| Field | Value |
|---|---|
| Severity | Low |
| Category | Correctness & logic bugs |
| Location | `src/MxGateway.Worker/Ipc/WorkerPipeClient.cs:169-171` |
| Status | Resolved |

**Description:** `retryAttempts` is computed as `(connectTimeout / min(connectTimeout, attemptTimeout)) - 1`. With defaults (30000 / 2000) this yields 14 retries, but each retry also incurs Polly exponential backoff. The overall `connectDeadline` (`CancelAfter(connectTimeout)`) is the real bound, so the computed attempt count can be larger or smaller than the time budget allows, and the formula is opaque.

**Recommendation:** Drive retries purely off the `connectDeadline` token (Polly stops when cancelled) and drop the fragile attempt-count arithmetic, or add a comment explaining the intent.

**Resolution:** 2026-05-18 — The opaque `retryAttempts` arithmetic in `ConnectWithRetryAsync` was removed. `MaxRetryAttempts` is now `int.MaxValue`, so the retry loop is bounded solely by the `connectDeadline` linked token (`CancelAfter(_connectTimeoutMilliseconds)`): Polly stops retrying the moment that token is cancelled, making the overall connect timeout the single source of truth and correctly accounting for the exponential backoff between attempts (which the old formula ignored). A comment documents the intent. No new test was added — the change does not alter observable behavior (the deadline was always the real bound; the old formula always permitted more attempts than fit the budget), and the existing `WorkerPipeClientTests.RunAsync_RetriesUntilPipeServerAppears` (server appears mid-retry) and `RunAsync_WhenPipeNeverAppears_ThrowsTimeoutException` (deadline ends the loop) already cover both retry-until-success and deadline-bounded termination.

### Worker-012

| Field | Value |
|---|---|
| Severity | Low |
| Category | Documentation & comments |
| Location | `src/MxGateway.Worker/MxAccess/MxAccessAlarmEventSink.cs:44-55`, `src/MxGateway.Worker/MxAccess/WnWrapAlarmConsumer.cs:38-43`, `src/MxGateway.Worker/MxAccess/MxAccessEventMapper.cs:106-112` |
| Status | Resolved |

**Description:** Multiple comments describe the alarm path as not-yet-wired future work ("PR A.2 — COM-side subscription scaffold … the worker advertises no alarm subscription", "the worker bootstrap will gain a thin 'run-on-STA' wrapper as part of A.3"). As of commit 6c64030 the alarm command handler, STA poll loop, and `SubscribeAlarms`/`AcknowledgeAlarm`/`QueryActiveAlarms` are all wired. These comments are stale and misleading.

**Recommendation:** Update the XML docs/comments to describe the shipped behavior; remove the "future PR" framing.

**Re-triage:** The `WnWrapAlarmConsumer.cs:38-43` citation is inaccurate — those lines were rewritten by Worker-001 and already describe the shipped no-internal-timer threading model correctly; nothing stale there. Conversely, two stale comments the finding did *not* cite were found on the same alarm path and fixed under the same root cause: `AlarmDispatcher.cs`'s `<remarks>` still framed the dispatcher as "the in-process slice of A.3" with a "companion follow-up PR" adding the (now-shipped) `SubscribeAlarmsCommand`/`AcknowledgeAlarmCommand`/`QueryActiveAlarmsCommand`, and stated the consumer "polls on a `System.Threading.Timer` thread today" — a claim made false by Worker-001's removal of that timer; and `AlarmCommandHandler.cs`'s `<remarks>` likewise asserted "the wnwrap consumer's polling timer fires on a thread-pool thread". The discovery document `docs/AlarmClientDiscovery.md` (referenced by the source comments) was deliberately left untouched: it is a historical research log of the investigation that chose the shipped design, not API/contract/lifecycle prose, and the source comments cite only its still-accurate "Option A — captured" payload schema.

**Resolution:** 2026-05-18 — Rewrote the stale alarm-path comments to describe shipped behavior with no "future PR / A.2 / A.3" framing. `MxAccessAlarmEventSink`: the class `<remarks>` and the `Attach` comment now explain that `AlarmDispatcher` owns the consumer→sink→queue wire-up and that `Attach` carries only the session id (no COM-event subscription is needed because the polled wnwrap consumer raises transition events itself). `MxAccessEventMapper.CreateOnAlarmTransition`'s XML summary now states the worker drives it from `MxAccessAlarmEventSink.EnqueueTransition` once `AlarmDispatcher` decodes a wnwrap transition. `AlarmDispatcher` and `AlarmCommandHandler` `<remarks>` were corrected to describe the shipped command surface and the no-internal-timer / STA-driven polling model (the `System.Threading.Timer` claims were factually wrong post-Worker-001). Pure documentation change — no behavior altered, no test needed; the build stays green.

### Worker-013

| Field | Value |
|---|---|
| Severity | Low |
| Category | Testing coverage |
| Location | `src/MxGateway.Worker/Sta/StaMessagePump.cs` |
| Status | Resolved |

**Description:** `StaMessagePump` — the heart of COM event delivery (`MsgWaitForMultipleObjectsEx` + `PeekMessage`/`DispatchMessage`) — has no direct unit tests. `StaRuntimeTests` exercises it indirectly for command wake-up but never verifies that a posted Windows message actually wakes the wait and is dispatched, nor that `PumpPendingMessages` returns a correct count. The alarm poll-loop lifecycle in `MxAccessStaSession` (start/cancel/await on shutdown) also has no test. These are the most failure-sensitive paths in the module.

**Recommendation:** Add tests that post a message to the STA thread and assert it is pumped, and tests covering alarm poll-loop start/stop and shutdown ordering.

**Re-triage:** This finding is stale as of the reviewed branch — the coverage it asks for already exists. `src/MxGateway.Worker.Tests/Sta/StaMessagePumpTests.cs` contains direct `StaMessagePump` tests covering null-argument validation, waking on a signalled event, returning on timeout, the zero-timeout conversion branch, `PumpPendingMessages` returning the correct count for messages posted to the STA thread (`PumpPendingMessages_MessagesPostedToStaThread_ReturnsCountProcessed`, `PumpPendingMessages_NoMessagesPosted_ReturnsZero`), and `WaitForWorkOrMessages` waking on a posted Windows message (`WaitForWorkOrMessages_WindowsMessagePosted_ReturnsForInputAvailable`) — exactly the "post a message and assert it is pumped" test the recommendation asks for. The alarm poll-loop lifecycle is covered by `MxAccessStaSessionTests.StartAsync_WithAlarmCommandHandlerFactory_PollOnceCalledViaSta` (start → poll runs on the STA) and `Dispose_StopsAlarmPollLoop` (Dispose joins the poll task; no further polls). The finding was raised against a stale view of the test project; no source or test change is required. Re-triaged as already resolved rather than fixed.

**Resolution:** 2026-05-18 — No code change. Re-triaged: the requested direct `StaMessagePump` tests (including posted-message dispatch and pump count) and the alarm poll-loop start/stop lifecycle tests already exist in `StaMessagePumpTests.cs` and `MxAccessStaSessionTests.cs`. See the re-triage note above for the specific test names.

### Worker-014

| Field | Value |
|---|---|
| Severity | Low |
| Category | Code organization & conventions |
| Location | `src/MxGateway.Worker/MxAccess/AlarmCommandHandler.cs:33`, `:202` |
| Status | Resolved |

**Description:** The file declares two public types — the `AlarmCommandHandler` class and the `IAlarmCommandHandler` interface. The C# style guide and the rest of the module follow one-public-type-per-file (e.g. interfaces in their own `I*.cs` files like `IMxAccessAlarmConsumer.cs`).

**Recommendation:** Move `IAlarmCommandHandler` to its own `IAlarmCommandHandler.cs` for consistency.

**Resolution:** 2026-05-18 — The `IAlarmCommandHandler` interface (with its XML docs) was moved verbatim out of `AlarmCommandHandler.cs` into a new `src/MxGateway.Worker/MxAccess/IAlarmCommandHandler.cs`, with its own `using` directives (`System`, `System.Collections.Generic`, `MxGateway.Contracts.Proto`). `AlarmCommandHandler.cs` now declares one public type, matching the module's one-public-type-per-file convention (cf. `IMxAccessAlarmConsumer.cs`). Pure file-organization change — no API surface, behavior, or namespace changed; no test needed. The worker build is clean with zero warnings (no unused usings left behind in `AlarmCommandHandler.cs`).

### Worker-015

| Field | Value |
|---|---|
| Severity | Low |
| Category | Correctness & logic bugs |
| Location | `src/MxGateway.Worker/MxAccess/MxAccessEventQueue.cs:115-145` |
| Status | Resolved |

**Description:** On overflow, `Enqueue` records the overflow fault and throws `MxAccessEventQueueOverflowException`; `MxAccessBaseEventSink.EnqueueEvent` catches it and calls `RecordFault` again. `RecordFault` is a no-op when a fault already exists, so the second call is harmless — but the intent is muddled, and there is no test asserting the dropped-event behavior. This is acceptable per the fail-fast design but undocumented at the call site.

**Recommendation:** Add a brief comment in `EnqueueEvent` clarifying that an overflow exception is expected and already self-records its fault, so the catch is intentionally a near no-op.

**Resolution:** 2026-05-18 — Added a comment in `MxAccessBaseEventSink.EnqueueEvent`'s catch block (per the finding's recommendation) explaining that two distinct fail-fast failures land there: a conversion failure from `createEvent()` (recorded here as an `MxaccessEventConversionFailed` fault) and an `MxAccessEventQueueOverflowException` from `Enqueue` at capacity, which — per the fail-fast backpressure design in `docs/DesignDecisions.md` — drops the event and has *already* self-recorded a `QueueOverflow` fault inside `Enqueue`. Because `MxAccessEventQueue.RecordFault` keeps only the first fault, the catch's `RecordFault` call is then a deliberate near no-op rather than a second, conflicting fault. Pure comment change as recommended — no behavior altered. `docs/DesignDecisions.md` already documents the fail-fast event backpressure rule, so no doc change was required.