mbproxy: Wave 4 — fix issues introduced by the Wave-1/2 fixes

Closes the new findings from the post-remediation re-review
(codereviews/2026-05-14/ReReviewAfterRemediation.md):

NC1 — ProxyWorker.StopAsync drain loop is structurally always-zero
  Wave 1's W1.5 inherited the original ShutdownCoordinator bug it was
  meant to replace. Supervisor.StopAsync nulls the per-mux counter
  provider before the drain loop runs, so CountInFlight always returns 0
  and the drain budget is never spent on actual draining. Fix: snapshot
  the in-flight count BEFORE supervisor stop, drop the theatrical
  post-stop loop, and report InFlightAtCancel as the snapshot count
  (= the number of in-flight requests dropped by the stop). The
  supervisor stop IS the drain — there is nothing to drain that
  wouldn't be killed by the stop itself.

NM1 — TearDownBackendAsync._connectGate.WaitAsync uncancellable
  Without a token, a long Polly-wrapped EnsureBackendConnectedAsync
  against an unreachable host could hold the gate for the full
  BackendConnectTimeoutMs * MaxAttempts window, blocking DisposeAsync
  (and therefore ProxyWorker.StopAsync) for that duration. Fix: bound
  the wait with a 2 s teardown deadline; on timeout proceed
  best-effort without the gate. Worst-case consequence is one orphaned
  in-flight cycle on the dying backend, surfaced to upstream as
  exception 0x0B by the watchdog.

NM2 — ReplaceContext non-atomic ctx + provider swap
  Snapshot path reads `_cacheStatsProvider` independently of `_ctx`. If
  `_ctx` was swapped first, a snapshot taken in the gap would still hold
  the OLD adapter wrapping the OLD cache — which the supervisor disposes
  immediately after we return. Fix: set the provider FIRST, then swap
  `_ctx`. Snapshots in the swap window now read either (old, old) or
  (new, new), never (old-after-disposed).

NM5 — Self-cascade ObjectDisposedException after dispose
  Writer/reader fault catches fired `_ = TearDownBackendAsync(...)`
  unconditionally. After DisposeAsync runs `_connectGate.Dispose()`, the
  fire-and-forget TearDown threw ObjectDisposedException on WaitAsync as
  an unobserved Task exception. Fix: skip self-cascade when
  `_disposeCts.IsCancellationRequested` — DisposeAsync runs an explicit
  TearDown anyway.

Nm1 — Saturation cleanup uses await SendResponseAsync
  W1.2's per-attacher delivery loop awaited the blocking SendResponseAsync,
  which would serialise on a wedged late-attacher's full bounded channel
  and stall delivery to its peers — contradicting the W1.3 doctrine that
  the fan-out path must never await per-pipe writes. Fix: use
  TrySendResponse and increment ResponseDropForFullUpstream on drop.

T2 — WatchdogVsResponse_Race seeded Random fragility
  Used `new Random(12345)` over [350, 450) ms with watchdog at 400 ms;
  Random's algorithm is implementation-defined across .NET major versions
  (legacy → Xoshiro128 in .NET 6) so a runtime upgrade could land all
  samples on one side of the deadline and break the "both branches must
  fire" assertion. Fix: deterministic counter-based alternation (15 fast
  + 15 slow across 30 iterations) — guaranteed by construction.

Latent items NM3 (_supervisorCts leak on re-Start) and NM4 (TCS
single-shot semantics) are unfixed: no caller actually re-Starts a
supervisor today; both become real only if the reconciler ever changes
to re-Start instead of dispose-and-rebuild. Documented in the re-review.

Tests: 387 pass / 0 fail. Three back-to-back race-test runs in
isolation all green (T2 alternation is deterministic).

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

mbproxy/codereviews/2026-05-14/ReReviewAfterRemediation.md

@@ -0,0 +1,126 @@
+# Re-Review After Remediation — 2026-05-14
+
+Re-review of the codebase after the six-commit remediation of the original 2026-05-14 review (Wave 1 → `ce32c5c`, Wave 2 → `e66b17f`, Wave 3 → `7ead358`, the easy 5 → `2545237`, the race-hard 5 → `53f842a`). Conducted via three parallel area-focused passes. **Eyes on what the fixes themselves introduced**, not what the original review already found.
+
+**Scope:** every src/ and tests/ change in `53a7111..HEAD` (37 files, ~+2000/−700 lines).
+
+## Headline
+
+The remediation is structurally sound. **One new critical finding** — the W1.5 drain loop is structurally always-zero — inherited the very bug it was meant to replace. **Five new major findings** cluster in two areas:
+
+- **W1.4 cascade gating** introduced two disposal-deadlock and exception-swallowing paths around the new `_connectGate.WaitAsync()` call inside `TearDownBackendAsync`.
+- **W1.1 + W1.5 lifecycle ordering** introduced a non-atomic context swap (snapshot reads can see a disposed cache during reseat) and silently single-shot supervisor semantics from the W2.15 + W2.16 combination.
+
+Plus a handful of test-discipline gaps from the new race-hard tests (reflection on private field names, RNG seed fragility across runtime versions).
+
+## New critical findings (1)
+
+### NC1. `ProxyWorker.StopAsync` drain loop is structurally always-zero — inherited the original `ShutdownCoordinator` bug it replaced
+**File:** `src/Mbproxy/Proxy/ProxyWorker.cs:276-310` (introduced by W1.5)
+
+**Path of execution:**
+1. `await Task.WhenAll(stopTasks)` — each `PlcListenerSupervisor.StopAsync` cancels its inner CTS, the listener's `RunAsync` exits, the OperationCanceledException catch calls `await listener.DisposeAsync()`, which calls `_multiplexer.DisposeAsync()`, which calls `_ctx.Counters.SetMultiplexProvider(null)` (`PlcMultiplexer.cs:224`).
+2. *Then* the drain loop begins. `CountInFlight` reads `supervisor.CurrentCounters.Snapshot().InFlightCount`, and `ProxyCounters.Snapshot()` (`ProxyCounters.cs:403-404`) returns `provider?.InFlightCount ?? 0`. The provider is now `null`. So `total == 0` on the first iteration and the loop exits immediately.
+
+The drain budget (`gracefulMs`) is never spent on actual draining. `inFlightAtCancel` will always log as `0`. The `mbproxy.shutdown.complete` event under load will report `InFlightAtCancel=0` even when in-flight requests were forcibly cancelled.
+
+**This is the same defect the original review (`AdminAndDiagnostics.md:C1`) called out in the deleted `ShutdownCoordinator`.** The W1.5 relocation faithfully copied the broken sequencing.
+
+**Fix options:**
+- (a) Snapshot in-flight counts BEFORE calling `supervisors.StopAsync`, then move drain-with-real-multiplexers into a phase BEFORE supervisor stop (i.e. stop accepting new connections, drain in-flight, then stop supervisors).
+- (b) Accept that supervisor stop *is* the drain (matches today's behavior) and remove the theatrical loop + `InFlightAtCancel` field from the log event.
+
+## New major findings (5)
+
+### NM1. `TearDownBackendAsync` acquires `_connectGate` with no token — disposal can be blocked indefinitely
+**File:** `src/Mbproxy/Proxy/Multiplexing/PlcMultiplexer.cs:333` (introduced by W1.4)
+
+`await _connectGate.WaitAsync().ConfigureAwait(false)` does not honour `_disposeCts.Token`. The dispose-driven teardown calls this after cancelling `_disposeCts`, but if a long Polly-wrapped `EnsureBackendConnectedAsync` is mid-connect against an unreachable host (waiting up to `BackendConnectTimeoutMs`), DisposeAsync blocks here for that duration. This blocks `ProxyWorker.StopAsync` past its drain budget.
+
+**Fix:** pass `_disposeCts.Token` (or a CTS that cancels on dispose) to `WaitAsync`, accept the `OperationCanceledException`, and skip teardown work if disposal already completed.
+
+### NM2. `ReplaceContext` writes `_ctx` and re-registers cache stats provider non-atomically — snapshot can see a disposed cache
+**File:** `src/Mbproxy/Proxy/Multiplexing/PlcMultiplexer.cs:173-184` (introduced by W1.1)
+
+Snapshot path (`ProxyCounters.Snapshot()` called from the admin endpoint) reads `_cacheStatsProvider` independently of `_ctx`. Sequence:
+1. Thread A enters `ReplaceContext`, writes `_ctx = newContext` (line 177).
+2. Thread B (admin) snapshots, reads `_cacheStatsProvider` (still the OLD adapter wrapping the OLD cache).
+3. Thread A runs `SetCacheStatsProvider(newCacheAdapter)`.
+4. Supervisor's `ReplaceContextAsync` proceeds to `oldCache.Clear()` then `oldCache.Dispose()`.
+
+A snapshot during the swap reports `cacheEntryCount`/`cacheBytes` from a cache the live multiplexer no longer references. Worse, if the snapshot races past step 4, the stats adapter holds a disposed `ResponseCache` and `Count`/`ApproximateBytes` may throw. **Same shape of race as the original review's "snapshot inconsistency" pattern**, re-opened on a new path by the W1.1 fix.
+
+**Fix:** swap the order (provider first, then `_ctx`) inside a short lock, OR have `ResponseCache.Count`/`ApproximateBytes` no-op on a disposed instance.
+
+### NM3. `_supervisorCts` leaks across `StartAsync` re-entry despite W2.16 guard
+**File:** `src/Mbproxy/Proxy/Supervision/PlcListenerSupervisor.cs:140, 437-446`
+
+W2.16's guard prevents re-Start while busy, but it does not address the leaking previous CTS the comment claims to fix. `StopAsync` sets `_state = Stopped` first, then awaits `_supervisorTask`. The previous `_supervisorCts` is disposed only in `DisposeAsync`, not in `StopAsync`. A re-Start pattern (Stop → Start without Dispose) leaks one CTS per cycle.
+
+Today no caller actually re-Starts a supervisor (ConfigReconciler's remove path uses `DisposeAsync`), so the leak is **latent**. Worth fixing now or documenting.
+
+### NM4. W2.15 TCS is never re-armed; W2.16 + TCS together make the supervisor effectively single-shot
+**File:** `src/Mbproxy/Proxy/Supervision/PlcListenerSupervisor.cs:74-75, 132-147`
+
+`_firstAttemptCompleted` is `readonly` and set in the field initialiser. Once `TrySetResult` fires (anywhere in `RunSupervisorAsync` including the `finally`), it stays completed forever. A restarted supervisor would have `WaitForInitialBindAttemptAsync` return immediately on the *previous* run's signal, regardless of the new run's bind status. Combined with the comment claiming "the supervisor's state machine has exactly one Start", the intent appears to be single-shot — but neither the guard nor the TCS enforces it cleanly.
+
+Today no caller re-Starts so this is **latent**. Either re-arm the TCS in `RunSupervisorAsync` or have W2.16 refuse re-Start unconditionally.
+
+### NM5. Self-cascade swallows `ObjectDisposedException` from `_connectGate` after disposal
+**File:** `src/Mbproxy/Proxy/Multiplexing/PlcMultiplexer.cs:451, 648` (introduced by W1.4 + W2.5)
+
+After `DisposeAsync` runs `_connectGate.Dispose()`, any in-flight self-cascade (`_ = TearDownBackendAsync(...)` from writer/reader fault paths) reaches `WaitAsync()` and throws `ObjectDisposedException`. The `try/finally` will then run `Release()` on a disposed semaphore in the `finally`, also throwing. Both throws happen on a fire-and-forget Task — they become unobserved exceptions on TPL's UnobservedTaskException event. Noisy but not fatal.
+
+**Fix:** wrap the body in `try { … } catch (ObjectDisposedException) { }` or short-circuit on `_disposed` before re-entering teardown.
+
+## New minor findings (8)
+
+| # | File:line | Finding |
+|---|-----------|---------|
+| Nm1 | `PlcMultiplexer.cs:842` (W1.2 cleanup) | Saturation cleanup uses `await SendResponseAsync` (blocking) per attached pipe. A single wedged late-attacher stalls saturation delivery to all others — contradicts the W1.3 doctrine. Use `TrySendResponse` here too. |
+| Nm2 | `PlcMultiplexer.cs:810` | W1.2 increments `CoalescedHit` for late attachers that ultimately receive exception 04. Counter shows hits that produced no real coalesced response. Document or decrement. |
+| Nm3 | `ProxyWorker.cs:269-310` | Both supervisor-stop and (theatrical) drain share `gracefulMs`. Worst-case shutdown is `2 * gracefulMs + 2s admin` — operators expect single budget. |
+| Nm4 | `PlcListenerSupervisor.cs:433-446` | `finally` writes `_state = Stopped` directly, preserving `_lastBindError` from a prior fault. After clean cancellation operators see "last bind error" that was never the cause of stopping. Doc the field as "last fault while running" or clear on clean exit. |
+| Nm5 | `EventLogBridge.cs:35-45` | W2.23 cached `_sourceExists` is correct, but no startup log confirms armed state. If `install.ps1` registers source after service starts, every Error+ event is silently dropped until restart. Add a one-line startup log. |
+| Nm6 | `ProxyWorker.cs:224-235` | W1.5 lazy `_admin = _services.GetService<AdminEndpointHost>()` returns `null` silently if `AddMbproxyAdmin()` is forgotten. Previous `IHostedService` registration would have errored loudly during host build. Use `GetRequiredService<>` or log a warning. |
+| Nm7 | `AdminEndpointHost.cs:200-210` | `DisposeAsync` lacks a `_disposed` guard; `ProxyWorker.StopAsync` calls `StopAsync` then DI disposes again. Operations are idempotent today but symmetry with `PlcMultiplexer` would be cleaner. |
+| Nm8 | `PlcMultiplexer.cs:415` | `TearDownBackendAsync` log `BackendDisconnected` event fires per cascade; queued cascades on the gate each log their own event with mostly-zero counts. Cosmetic noise. |
+
+## Test-discipline findings (4)
+
+| # | File:line | Finding |
+|---|-----------|---------|
+| T1 | `PlcMultiplexerTests.cs:1152-1165`, `SupervisorTests.cs:1779-1789` | `DrainAllocator` and the runtime-fault test reflect on private field names (`_allocator`, `_currentListener`, `_listener`). A rename refactor breaks them at run-time, not compile-time. Recommend `[InternalsVisibleTo]` test seam OR comments on each reflected field warning that tests depend on the name. |
+| T2 | `PlcMultiplexerTests.cs:1379-1380` (W3 #8) | `WatchdogVsResponse_Race` uses `new Random(12345)` over `[350, 450)` ms with watchdog at 400 ms; asserts BOTH branches hit. `Random` with a seed is implementation-defined across major .NET versions (legacy → Xoshiro128 in .NET 6). On a runtime upgrade the seed could land all 30 samples on one side and the test breaks. Replace with deterministic alternation: `int delay = (i & 1) == 0 ? 350 : 450;`. |
+| T3 | `Mbproxy.Tests.csproj:38-40` (W2.21) | `RemoveInheritedAppsettings` Target uses `AfterTargets="Build"` and deletes from `$(OutputPath)`. Won't fire on `dotnet publish`. Test projects rarely publish but worth `AfterTargets="Build;Publish"` + a second Delete against `$(PublishDir)appsettings.json`. |
+| T4 | `InFlightByKeyMapTests.cs:46, 75, 112` | After W3 removed the `bool` return, test method names still read `TryAttachOrCreate_..._ReturnsTrue_WasNewTrue`. Decorative but misleading for grep. |
+
+## Verified clean (sampled, not exhaustive)
+
+- **W2.3 ConcurrentDictionary migration** on `_supervisors` — all mutations atomic; status-page enumeration lock-free; Restart's "remove + add" two-step is per-key (parallel keys disjoint by name). 
+- **W2.1 coalescingAccessor propagation** — both Add and Restart paths receive it; Reseat correctly does not (same supervisor, same multiplexer, same accessor).
+- **W2.13 OOR check** — multiplication is bounded by the guard; even worst-case `9999 * 10_000 + 9999 = 99_989_999` fits in int32 without overflow.
+- **W2.14 byteCount validation** — strict `<` check passes a perfectly-sized PDU; trailing-byte case correct.
+- **W2.10 resolved-TTL re-check** — `BcdTagMapBuilder.Build` is called exactly once per PLC at validation; no duplicate work.
+- **W2.18 ConnectionOptions validation** — both `MbproxyOptionsValidator` and `ReloadValidator` reject `<= 0`; no bypass path.
+- **W3 `HasBadNibble` dedupe** — clean; the codec's internal helper is the single source of truth.
+- **W2.15 TCS signalled in every exit path** of `RunSupervisorAsync` (bind-success, bind-failure-into-recovery, run-failure-into-recovery, finally) — no hang on `WaitForInitialBindAttemptAsync` for the first run.
+- **W2.17 TransitionTo lock contract** — both writers use it; `Snapshot` reads under the same lock; no torn triples.
+- **`TxIdAllocator.Release` double-call is benign** (`TxIdAllocator.cs:121-129` checks `if (_inUse[id])`); the W1.4 channel drain releasing a TxId already released by the correlation drain is safe.
+- **W1.1 in-PDU snapshot consistency** — `OnUpstreamFrameAsync` reads `_ctx.Cache` and `_ctx.TagMap.ResolveCacheTtlMs` non-atomically; the only mid-PDU swap visible would change cache eligibility, not produce corrupted output. Downstream `WithCurrentRequest` snapshots TagMap+Cache for the rewriter, so the rewrite itself is consistent.
+- **W2.7 cache-FC byte sourced from post-rewriter buffer** — correct; the rewriter never touches the FC byte but the source must remain `frame[…]` to capture the exception bit.
+
+## Recommended next actions
+
+If you want to keep the issue tracker tidy, these are the items that should land as a small follow-up commit:
+
+1. **NC1** — fix or remove the theatrical drain loop. Smallest change: delete the loop and the `InFlightAtCancel` field from the log.
+2. **NM1** — pass `_disposeCts.Token` to `_connectGate.WaitAsync` in `TearDownBackendAsync`.
+3. **NM2** — wrap the `_ctx` write + provider re-registration in a single short lock, or null-guard the cache adapter against disposed-cache reads.
+4. **NM5** — wrap the self-cascade body in `try { … } catch (ObjectDisposedException) { }`.
+5. **Nm1** — use `TrySendResponse` in the W1.2 saturation cleanup loop to avoid wedge-by-wedged-attacher.
+6. **T2** — replace the seeded `Random` in `WatchdogVsResponse_Race` with deterministic alternation.
+
+NM3 and NM4 are latent (no current caller exercises supervisor re-Start). They become real only if the reconciler ever changes to re-Start instead of dispose-and-rebuild. Defer or document.
+
+Everything else is documentation / cosmetic / minor cleanup.