Stabilize unit tests (#932, #938)

[jhugard]

Problem (#932, #938)

Unit tests for async operations and task queues exhibit intermittent failures on win32 target (others untested) due to timing assumptions, races, and use-after-free in test fixtures (#932). In addition, AsyncBlockTests::VerifyAsyncBlockReuse can crash under stress when concurrent async lifecycle phases append to a shared std::vector (#938).

Key failure modes:

• Flaky async/task-queue unit tests due to timing assumptions and use-after-free bugs #932 UAF: VerifyDuplicateQueueHandle tests an error path by passing a freed queue handle, which can be dereferenced in the duplicate-handle failure path.
• Data Race in AsyncBlockTests::VerifyAsyncBlockReuse #938 race: cleanup of the first async call can run concurrently with begin of the second call, leading to simultaneous opCodes.push_back() and heap corruption during vector reallocation.

Solution

Stabilize async/task-queue unit tests and harden factorial worker lifetime (#932)

• Explicit synchronization: Add wait loops in VerifySimpleAsyncCall, VerifyWaitForCompletion, and VerifyDistributedAsyncCall to ensure async operations complete before verification.
• Queue drains: Wait for both Work and Completion ports to be empty before verifying final state.
• Relaxed timing: Allow per-iteration slack window for VerifyDistributedAsyncCall timing check (5 iterations x 100ms with one-interval slack).
• Lifetime hardening: Hold per-call references in FactorialWorkerSimple and FactorialWorkerDistributed to prevent use-after-free during async callbacks.
• Conditional waits: In _VerifyQueueTermination, only wait for counts to settle for non-blocking termination cases.

Fix use-after-free in test VerifyDuplicateQueueHandle (#932)

• Replace freed queue handle with stack-allocated fake handle (64-byte aligned buffer) for error testing.
• Prevents dereferencing of previously-freed memory in the XTaskQueueDuplicateHandle failure path.

Fix data race in AsyncBlockTests opcode logging (#938)

• Replace std::vector<XAsyncOp> with a fixed-capacity lock-free log: std::array<std::atomic<XAsyncOp>, 16> + std::atomic<size_t>.
• Use store(memory_order_release) for writes and load(memory_order_acquire) for snapshots to ensure visibility on weakly-ordered architectures (ARM).
• Provide RecordOp() for lock-free append and GetOpCodes() for snapshots.
• Snapshot optimization for repeated verification checks.

Impact

• Eliminates timing-dependent test failures and data races (Flaky async/task-queue unit tests due to timing assumptions and use-after-free bugs #932, Data Race in AsyncBlockTests::VerifyAsyncBlockReuse #938).
• Prevents use-after-free in test fixtures.
• Improves test reliability on high-load systems and CI.
• No impact on product code or API.

Testing

• VerifyAsyncBlockReuse: 10/10 passes.
• AsyncBlockTests suite: 23/23 passes.
• Note: The original data-race crash required a long soak with page heap; extended stress testing is recommended for maximum confidence.

[brianpepin]

These changes look good to me; thanks for contributing.

[jhugard]

Sneaking in another change to relax a time-based test, found while soaking under cdb with page-heap. I've found that a loaded system sometimes needs >50ms epsilon to service delayed tasks.

[jhugard]

Don't merge this PR just yet. There was an independent latent bug #938 and the fix obsoletes the need for a change in this PR.

[jhugard]

Clean and ready for review.