fix(#2809): the two IoPoolTest gates use the sanctioned bridge, and release in a finally by rbuergi · Pull Request #2822 · Systemorph/MeshWeaver

rbuergi · 2026-08-30T18:28:12Z

Two TaskCompletionSource-as-a-gate sites in test/MeshWeaver.Hosting.Test/IoPoolTest.cs — the shape AGENTS.md names in the same breath as ManualResetEventSlim. Found by the Copilot review on #2806 and deliberately left out of it, whose scope was the 79 measured ManualResetEventSlim sites plus the guard's move to zero-tolerance.

Why the ratchet cannot catch these

HandWovenGateRatchetGuard matches the primitives that park a thread by name (SemaphoreSlim | ManualResetEventSlim | ManualResetEvent | AutoResetEvent | CountdownEvent | Monitor.Wait). A TaskCompletionSource used as a gate is structurally the same thing, but the type has a hundred legitimate uses — a name match would cry wolf on all of them. So these needed a human finding, and the fix is worth landing rather than waiting for a guard to grow a rule it cannot express.

The two sites

site	note
`Invoke_caps_in_flight_at_the_pool_bound`	20 pooled leaves parked on one signal, created without `RunContinuationsAsynchronously` — a single `SetResult()` resumed every waiting leaf inline on the test thread, the exact inline-resumption defect `ObservableToTaskBridgeGuard` exists for
`Drain_joinsABlockingLeaf_evenWhileAnAsyncLeafIsRunning`	the async leaf's gate; carried the flag, so the milder of the two

Neither released in a finally. An assertion that threw before the release left the parked leaves waiting on a signal that never arrived — so the host hung, and an ordinary assertion failure with a perfectly good message was reported as a timeout.

The fix

AsyncSubject<T> + ObservableAwait.Await() — the one sanctioned bridge, which sets RunContinuationsAsynchronously itself — with the release in a finally at both sites. Completing twice is a no-op, so the happy path and the finally can both release without a flag.

No cancellation token on the asyncGate Await(), deliberately: that test's premise is that the async leaf is still in flight when the blocking join runs, so a ct-aware wait unwinding on the drain's cancel would weaken exactly what it measures.

Mutation-checked on the property that matters

Made an assertion fail before the release (Be(cap) → Be(cap + 99)):

MeshWeaver.Reactive.Assertions.AssertionException : Expected value to be 102, but found 3.
Failed!  - Failed: 1, Duration: 22 ms

22 ms and the real message — not a hang. Without the finally, that same mutation strands 20 pooled leaves and the assertion is reported as a timeout.

dotnet build test/MeshWeaver.Hosting.Test -c Release -warnaserror        0 Error(s)
dotnet test  test/MeshWeaver.Hosting.Test  (whole project)   269 passed, 0 failed, 9 s

Fixes #2809

🤖 Generated with Claude Code

https://claude.ai/code/session_014FhP7Jg65q8cAATnFfeaTK

…elease in a finally test/MeshWeaver.Hosting.Test/IoPoolTest.cs held two TaskCompletionSource-as-a-gate sites — the shape AGENTS.md names in the same breath as ManualResetEventSlim. Found by the Copilot review on #2806 and deliberately left out of that PR, whose scope was the 79 measured ManualResetEventSlim sites plus the guard's move to zero-tolerance. ## Why the ratchet cannot catch them HandWovenGateRatchetGuard matches the primitives that park a thread BY NAME (SemaphoreSlim | ManualResetEventSlim | ManualResetEvent | AutoResetEvent | CountdownEvent | Monitor.Wait). A TaskCompletionSource used as a gate is structurally the same thing, but the type has a hundred legitimate uses, so a name match would cry wolf. That is why these two needed a human finding rather than a ratchet, and why the fix is worth landing rather than waiting for the guard to grow a rule it cannot express. ## The two sites * `Invoke_caps_in_flight_at_the_pool_bound` — 20 pooled leaves parked on one signal, created WITHOUT RunContinuationsAsynchronously. A single SetResult() resumed every waiting leaf INLINE on the test thread: the exact inline- resumption defect ObservableToTaskBridgeGuard exists for. * `Drain_joinsABlockingLeaf_evenWhileAnAsyncLeafIsRunning` — the async leaf's gate. Carried the flag, so the milder of the two. Neither released in a `finally`: an assertion that threw before the release left the parked leaves waiting on a signal that never arrived, so the host HUNG and an ordinary assertion failure was reported as a timeout. ## The fix AsyncSubject<T> + ObservableAwait.Await() — the one sanctioned bridge, which sets RunContinuationsAsynchronously itself — with the release in a `finally` at both sites. Completing twice is a no-op, so the happy path and the finally can both release without a flag. No cancellation token on the asyncGate Await, deliberately: that test's premise is that the async leaf is STILL in flight when the blocking join runs, so a ct-aware wait unwinding on the drain's cancel would weaken exactly what it measures. ## Mutation-checked, on the property that matters Made an assertion fail BEFORE the release (`Be(cap)` → `Be(cap + 99)`): MeshWeaver.Reactive.Assertions.AssertionException : Expected value to be 102, but found 3. Failed: 1, Duration: 22 ms 22 ms and the real message — not a hang. Without the finally that same mutation strands 20 pooled leaves and the assertion is reported as a timeout. dotnet build test/MeshWeaver.Hosting.Test -c Release -warnaserror 0 Error(s) dotnet test test/MeshWeaver.Hosting.Test (whole project) 269 passed, 0 failed Fixes #2809 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FhP7Jg65q8cAATnFfeaTK

Copilot

Pull request overview

This PR fixes two remaining “TaskCompletionSource-as-a-gate” patterns in IoPoolTest by switching to the sanctioned AsyncSubject<T> + ObservableAwait.Await() bridge and ensuring the release happens from a finally, so assertion failures don’t strand pooled workers and masquerade as timeouts.

Changes:

Replace TaskCompletionSource gate usage with AsyncSubject<T> plus ObservableAwait.Await() in two IoPoolTest cases.
Move gate release to finally blocks to prevent hangs when assertions fail before the release path.
Add a What’s New entry documenting the user-noticeable impact (tests reporting the real failure instead of hanging).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
test/MeshWeaver.Hosting.Test/IoPoolTest.cs	Replaces the two TCS gate sites with `AsyncSubject` + `Await()` and guarantees release in `finally`.
src/MeshWeaver.Documentation/Data/WhatsNew/2026-08-30-two-test-gates-stop-resuming-inline.md	Adds a What’s New entry describing the fix and its effect on test failure reporting.

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

github-actions · 2026-08-30T18:34:49Z