test(CM0001): observe a cancelled request's disconnect off the thread pool; overwrite test-result artifacts on re-run

[Arthurvdv]

Summary

Fixes the flaky CancelledCompilation_ClosesHttpRequestPromptly_AndNextCompilationRecovers test in ALCops.Common.Test and the stale "Test results" check after a single-leg re-run. Both surfaced on #571 (run 36340330338), which does not touch Common; the failure is unrelated to that PR.

What happened

• Attempt 1, leg 12.1.13.35966: the test took exactly 2 s and only closedPromptly failed. next was empty and RequestCount == 2 passed, so cancellation had propagated and nothing was cached. The 2 s Task.Delay race won on a slow shared runner.
• The re-run (attempt 2) passed the test in 10 ms, yet the check still showed the failure and listed every trx twice: the run held two test-results-12.1.13.35966 and two test-results artifacts (one per attempt). download-artifact picked up the attempt-1 per-version artifact, so even the attempt-2 unified artifact contained the failing trx, and dorny downloaded both unified artifacts.

Mechanism

Cancel() runs the linked-token registrations inline (test token → SDK linked CTS → HttpClient linked CTS → HttpConnection.Dispose), so the client socket is closed before Cancel() returns. The fake server observed that with ReadAsync plus a RunContinuationsAsynchronously completion feeding Task.WhenAny, which needs more thread-pool hops than the timer it raced. On a saturated pool the timer wins although the close was immediate.

Changes

• Test: the stalled request's disconnect is observed with a blocking NetworkStream.Read on a LongRunning task that sets a ManualResetEventSlim; the test waits on that signal. The 2 s window and every assertion are unchanged. The blocking read is bounded by the server's shutdown budget because the shutdown token cannot interrupt it.
• CI: overwrite: true on the per-version and unified test-result uploads in build-test.yml, so a re-run replaces the same-named artifact.
• Docs: testing.md (Concurrent synchronous HTTP tests), the CM0001 rule doc test notes, and release-strategy.md (Test report gate) record the why.

Verification

• Best-effort local repro of the original test under DOTNET_ThreadPool_MinThreads=2 / MaxThreads=3: 20/20 passed, so the flake did not reproduce on this Windows machine. The fix rests on the mechanism above and the attempt-1 log evidence.
• Fixed test under the same pressure: 20/20 passed. Whole ALCopsSettingsRemoteRecoveryTests fixture without pressure: 10/10 runs passed. Full ALCops.Common.Test suite: 159 passed.
• dotnet format ALCops.sln --verify-no-changes clean; Validate-Rules.ps1 OK.
• The re-run path can be checked on this PR's own run: re-run one test leg and confirm the artifact list holds one artifact per name and the check lists each trx once.

🤖 Generated with Claude Code