[Copilot]

Follow-up to krwq's request, addressing the blocked V2 cleanup worker and premature asynchronous cleanup.

• Complete single-event callback removal and wait-handle disposal before returning, without a detached cleanup task. Dispatch user callbacks outside the subscription lock and synchronize signaling with disposal to avoid cancellation-and-join deadlocks and late-callback races.
• Use the base asynchronous wrapper around the Raspberry Pi synchronous wait, retaining pin ownership and the linked cancellation-token source until completion.
• Stabilize the BMP3xx simulator tests that were intermittently failing macOS CI. Three timing-window tests now use FakeTimeProvider instead of real-time scheduling; the oversampling regression checks rejection at 20 ms and acceptance at 68 ms. The simulator defaults to system time, and BMP3xx production code is unchanged.

Validation on Windows with the actual .NET 8/xUnit runner: all 89 BMP3xx tests passed in both Debug and Release; all four Windows-supported GPIO lifetime cases passed in both configurations using the repository's existing platform filter. The reviewed timing repair also passed 50 repeated simulator-suite runs (450 test executions) and a controlled real 100 ms scheduling delay. All eight GPIO lifetime cases also passed on Ubuntu with the actual .NET 8/xUnit runner in both Debug and Release, including all four Unix-only Raspberry Pi cases. No hardware testing is claimed.

All six Debug/Release matrix jobs across Linux, macOS, and Windows passed in cross-platform CI on the combined commit 39d244de0b230d211c67f874bd5db7d91c883c7d.

Concurrent-wait reference counting, callback removal by event type, and EdgeEvent disposal remain outside this change.

[krwq]

/azp run dotnet.iot

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).

[krwq]

@copilot Please fix a cancellation deadlock introduced by this PR and update this same PR. In LibGpiodV2EventObserver, the wait's finally now acquires _handlersBySubscription before the wait task completes (lines 47–57). HandleEdgeEvent holds that same lock while invoking user callbacks (lines 246–272). Concrete scenario using LibGpiodV2Driver directly: register an ordinary callback on one offset, start a single asynchronous wait on a DIFFERENT offset, and have the callback cancel that wait and synchronously join its task. The subscription-registry lock is shared across offsets, so this does not require concurrent operations on the same pin. The callback waits for the task; the task's cleanup waits for the callback-held lock. Neither can complete. This is introduced by the new synchronous cleanup, not the separate reference-counting finding. Preserve deterministic cleanup before wait completion without invoking arbitrary user callbacks under the subscription-registry lock. Make callback dispatch versus wait-handle removal/disposal safe as well: merely moving callback invocation outside the lock can otherwise invoke an already-disposed wait handle. Do not restore a detached cleanup task to hide the deadlock. Add a bounded, hardware-independent cancellation-and-join regression and coverage of callback dispatch racing with wait cleanup. Run the targeted GPIO tests and an independent fresh-reader review before finalizing. Commit the correction to this PR's existing branch; do not open another PR or force-push. Keep the independently merged BMP3xx timing failure separate; its local repair is already being investigated. The unrelated callback reference-counting/event-type-removal and EdgeEvent-disposal changes remain outside this PR.

[Copilot]

@copilot Please fix a cancellation deadlock introduced by this PR and update this same PR. In LibGpiodV2EventObserver, the wait's finally now acquires _handlersBySubscription before the wait ta...

Fixed in adc11cc on this branch. Callback snapshots are dispatched outside the registry lock; wait-handle signaling and disposal are synchronized so captured callbacks are safe after cleanup. Added bounded, repeated cancellation-and-join and dispatch/cleanup regressions. Fresh independent review found no significant issues. GPIO build and targeted tests were attempted but blocked before compilation by Arcade feed DNS failure; CodeQL timed out. BMP3xx and the other excluded changes remain untouched.