Restore green CI on main #21246

Investigation notes

• Explicit-start lifecycle / Windows: Failures span startup/readiness and shutdown. The current run fails waiting for the resource predicate, not in StopAsync. Next: Investigate the current startup/readiness failure separately. PR Add explicit-start shutdown regression coverage (startup-token cancellation and process exit) #21397 addresses shutdown-token ownership, not every failure of this method.
• Azure sandbox cancellation: Assert.False: expected false, actual true. Next: Investigate cancellation ordering.
• Extension diagnostics lock holder: Storage diagnostics' exclusively held lock test loses its lock-holder process. Next: Investigate lock-holder lifecycle.
• Extension resource debugger control: Packaged debugger attach/breakpoint/detach test fails applying an E2E control revision. Next: Investigate debugger/control lifecycle; do not merge with runner-loss occurrences in this shard.
• Azure Functions shutdown: Multi-project test is cancelled during application shutdown. Next: Investigate shutdown separately from the explicit-start test's fix.
• Terminal-host socket cleanup: Socket-cleanup test throws cancellation on Windows. Next: Investigate shutdown and socket cleanup.
• Deno startup / telemetry fixtures: Functional-fixture timeouts and telemetry initialization/address failures are grouped by area; they are not proven to share one root cause. Next: Investigate the fixture/telemetry variants separately. The linked issues track the two functional-fixture tests.
• DevTunnel disposed service provider: Shared observer throws ObjectDisposedException for IServiceProvider on Windows and Ubuntu. Next: Investigate shared-observer lifetime.
• MongoDB replica-set startup: Resource rs0 fails to start or become healthy across healthy-node, topology/data-reuse and Mongo Express test variants. These share the observed startup failure, not a proven underlying cause. Next: Investigate replica-set startup.
• Rebuild-command assertions: Command-success and build-count assertions fail in the same test. Next: Keep Assert.True and build-count mismatch signatures distinct.
• Terminal-host owner exit: The managed terminal-host owner-exit test fails. Next: Compare the current exception/exit details with the tracked owner-exit timeout.
• Dashboard exemplar dialog: Assert.Single finds an empty exemplar-dialog collection. Next: Investigate dialog readiness and exemplar availability.
• Terminal-host RPC cancellation: Cancellation-delivery test reaches its 30-second timeout on Ubuntu. Next: Investigate cancellation delivery.
• Java container telemetry assertion: Telemetry URL assertion fails. The latest occurrence expects http://:4317 but receives http://. Next: Investigate the URL construction/configuration signature; do not assume a past hostname-assertion fix covers this.
• Extension Deno debug startup: Debug AppHost startup times out; teardown also encounters EBUSY. Next: Investigate startup and cleanup; do not assume an infrastructure flake.
• Integration-host registration: Registration/discovery test times out on Windows. Next: Investigate registration/discovery ordering.
• CLI ambient-discovery timeout: The macOS command-resolution test times out after 30 seconds. Next: Investigate command/test lifetime; do not conflate this assertion with a later test-host exit.
• Extension Node Functions breakpoint: Same E2E test times out waiting for the startup breakpoint in Program.cs:4. Next: Investigate the startup breakpoint/control path; do not infer an HTTP-request breakpoint failure.
• CLI graceful-stop escalation: Both console-isolation variants hit the same 30-second timeout while testing escalation when a process ignores graceful stop. Next: Investigate the process wait/termination path; compare the two theory variants.

• Windows runner loss — investigate: Explicit runner-communication-loss annotations across different Windows .NET suites and extension E2E shards. Next: Correlate runner/process/resource diagnostics. The closed historical issue is not proof of today's cause.
• Homebrew local-tap trust — fix merged; verify newer image: The cask audit rejects the untrusted local tap. Next: Recorded successful trust/audit evidence used image 20261021; verify the affected newer image before closing out.
• Windows post-test process startup — investigate: Post-test processes fail with exit -1073741502 after a passing test summary. Next: Investigate process initialization rather than test assertions.
• Artifact-service errors — investigate: Artifact listing/creation timeouts and a busy-server upload failure. These are service/transport symptoms, not one proven cause. Next: Separate service failures from assertions and runner loss.
• macOS CLI test-host exit 137 — investigate: Test application exits 137; the outer test command exits 7. Some jobs also contain separate failed tests. Next: Investigate host/session termination. Exit 137 alone does not establish OOM.

Other recorded failures

Separate signatures can occur in the same suite; these rows are not additional recurring-pattern counts.

Daily runner-loss counts

No matching failure was observed in the last three weekdays. A quiet stretch is not proof of a fix or of sufficient later executions. Zero weekly counts can mean the last observation is outside this window.

Verified recoveries

• Historical Go includeDocs compilation break: newer package compilation succeeded: successful build evidence (2026-10-30).

Coverage and counting

• Run-start window: 2026-10-23T20:24:02Z through 2026-10-30T20:24:02Z. Run inventories missing: 0; failed-job evidence available: 244/245.
• Weekdays use run-start UTC date; last seen uses failed-job completion time. Monday uses Thursday, Friday and Monday. Weekend failures remain in weekly/current evidence.
• Counts include failures observed before workflow cancellation and in still-running workflows. Cancellation alone is not a failure, and missing results are not passes.
• Affected runs are deduplicated per reason. Groups overlap; do not add their affected-run totals. Execution eligibility has not been established, so no failure-rate denominator is claimed.
• Current observations use the newest run and latest completed non-cancelled run, at their latest attempts. Older failures are not automatically carried forward as current.
• Merged fixes, closed issues, later green jobs and quiet periods do not establish permanent resolution. This is a CI-blocker report, not an exhaustive flaky-test tracker.
• Logs/annotations and retained local observations supply counts; issue/PR bodies supply only verified tracking context.