test: flaky LT_OnMountUnmountBalanced selftest — unmount count asserted before async OnUnmount flush completes

[azchohfi]

Summary

Intermittent selftest failure: LT_OnMountUnmountBalanced (asserts LT_Unmounts_Exactly_30 + LT_NoLeak_MountsEqualUnmounts) in tests/Reactor.AppTests.Host/SelfTest/Fixtures/LifecycleTortureFixtures.cs. A generic lifecycle-torture fixture that mount/unmount-counts a TextBlock across rapid toggles.

Root cause (test-side race, not product code)

LT_Mounts_Exactly_30 passes (mounts == 30), but the final drive!(0) clear batch's OnUnmount callbacks have not all flushed when the unmount count is asserted → unmounts < 30 → assertion fails. The exactly-conserved unmount count is checked immediately after the clear batch rather than after a deterministic quiesce of the async OnUnmount dispatch.

Evidence

• Observed on PR perf: cache per-render arrays/LINQ in DataGrid #669 CI, commit 9f640ba0, Selftests job run 28200113984: 1/1209 failed (only LT_OnMountUnmountBalanced).
• gh run rerun 28200113984 --failed → Selftests PASSED (7m15s) on the same commit, zero code change → confirmed non-deterministic.
• Not caused by perf: cache per-render arrays/LINQ in DataGrid #669: that branch touches only 3 DataGrid files (no AppTests.Host changes; only static readonly immutable arrays + pure methods). Every DataGrid fixture passed; the failing fixture doesn't use DataGrid. Fixture last modified by unrelated commit 222a299a.

Impact

False-red on the Selftests CI job that can hit any PR intermittently. Relevant right now because the perf-improvement fleet is about to go through an individual per-PR gh pr ready + CI + /perf GO cycle — a flake-induced red Selftests run during GO must not be mistaken for a real regression.

Suggested fix direction

Make the unmount-count assertion await a deterministic OnUnmount flush (drain/quiesce the dispatcher so all OnUnmount callbacks from the final drive!(0) clear batch have run) before asserting the conserved count — rather than asserting immediately after the clear batch. Alternatively add an explicit sync point that guarantees all OnUnmount callbacks completed.

Scope note

Not part of the /perf self-contained harness build fix (#677) — that is narrowly the StressPerf harness build, not selftest stability. Tracking this separately so it can be fixed independently / post-fleet.

Refs #669. Flagged by the Datagrid-allocs perf session.

[azchohfi]

Second independent occurrence — confirms this is cross-PR, not introduced by any one change.

• PR perf: eliminate per-diff allocations in ChildReconciler + KeyedListDiff #657, commit 83599a0, Selftests run 28200086413: the same single fixture LT_OnMountUnmountBalanced failed (LT_Unmounts_Exactly_30 / LT_NoLeak_MountsEqualUnmounts; unmounts != 30). All other 16 checks passed. gh run rerun 28200086413 --failed then went green with zero code change.
• That PR production delta is unrelated to lifecycle: a boolPool.Return clearArray flag in keyed-list-diff; both buffers self-clear before read, so counts cannot change. The fixture uses distinctly-keyed TextBlocks with no key overlap (pure remove+mount, never hits the docs(skills): split SKILL.md into root + on-demand sub-skills #30 skip path). Local run 6/6 pass on its HEAD.
• Now reproduced on TWO unrelated fleet PRs (perf: cache per-render arrays/LINQ in DataGrid #669 and perf: eliminate per-diff allocations in ChildReconciler + KeyedListDiff #657) with different small diffs, both green on re-run. Strengthens the conclusion: pre-existing intermittent harness-timing flake in the async real-WinUI selftest, not a regression.