Found while investigating a different (now-closed) flake on current main 77e56cb0.
Evidence
12 full @pretable/react suite runs (1646 tests each) executed under synthetic load (1-min load 89–156 on a 10-core Mac). 5 of the 12 runs failed, all in the eviction-* files, none anywhere else:
eviction-population-change.test.tsx > … > paints nothing rather than rows the reader never selected — ×3eviction-gate-blip.test.tsx — three distinct tests, ×1 eacheviction-selection-surface.test.tsx > … > survives an incremental slide that evicts one endpoint, then extends from it — ×1
The same harness executed the previously-suspected external-filter-authority test 72 times (60 isolated + 12 in-suite) at those loads with zero failures, so the load rig demonstrably produces real failures and these are where they land.
Why it is worth a look
eviction-population-change failing three times on the same assertion is a cluster, not scattered noise. Worth determining whether these are timing-sensitive assertions that should poll for the settle (the house idiom — see the { timeout: 15_000 } / expect.poll(…, { timeout: 20_000 }) sites) or a genuine eviction/selection race that only starvation exposes.
Method that worked: run the full react suite 2–3× concurrently plus CPU spinners, loop, and capture each run's output to its own file (never pipe a gate through grep|head — SIGPIPE truncates it).
🤖 Generated with Claude Code
Found while investigating a different (now-closed) flake on current main
77e56cb0.Evidence
12 full
@pretable/reactsuite runs (1646 tests each) executed under synthetic load (1-min load 89–156 on a 10-core Mac). 5 of the 12 runs failed, all in theeviction-*files, none anywhere else:eviction-population-change.test.tsx > … > paints nothing rather than rows the reader never selected— ×3eviction-gate-blip.test.tsx— three distinct tests, ×1 eacheviction-selection-surface.test.tsx > … > survives an incremental slide that evicts one endpoint, then extends from it— ×1The same harness executed the previously-suspected
external-filter-authoritytest 72 times (60 isolated + 12 in-suite) at those loads with zero failures, so the load rig demonstrably produces real failures and these are where they land.Why it is worth a look
eviction-population-changefailing three times on the same assertion is a cluster, not scattered noise. Worth determining whether these are timing-sensitive assertions that should poll for the settle (the house idiom — see the{ timeout: 15_000 }/expect.poll(…, { timeout: 20_000 })sites) or a genuine eviction/selection race that only starvation exposes.Method that worked: run the full react suite 2–3× concurrently plus CPU spinners, loop, and capture each run's output to its own file (never pipe a gate through
grep|head— SIGPIPE truncates it).🤖 Generated with Claude Code