PG Tests baseline flake rate ≈1/3626 per run — surfaced by wish pg-test-perf

[namastex888]

Summary

After landing wish pg-test-perf, PG Tests (serial, 208 files, 3626 tests) consistently shows 1-3 flaky failures per run, with the failing test different every time. Across 7 consecutive CI runs on PR #1317:

Run	Wall-clock	Flakes	Which tests
1	145s (pre-timeout-bump)	3	events-stream drain, qa-runner × 2
2	64.9s	2	Team Manager createTeam, turn-close happy path
3	67.5s	1	(unnamed) beforeEach timeout
4	87.9s	2	serve lifecycle, events-stream cursor
5	64.9s	1	Group 1 observability migrations
6	76.1s	1	pg > register and get
7	68.7s	1 + 2 errors	pg > syncWishes upserts

Never the same test twice. That's the classic signature of environmental noise — Blacksmith runner jitter, pgserve cold-start variance, NOTIFY/LISTEN timing, fixture-order dependence. Not a deterministic code bug.

Impact

• Merge velocity: every PR needs a gh run rerun --failed lottery to hit green.
• Signal erosion: engineers learn to ignore "1 fail" as noise. Real regressions hide in the noise.

Proposed approaches (ranked by ROI)

1. Retry-failed-tests-once inside bun test (custom reporter or wrapper) — masks intermittents, surfaces consistent fails. The fastest ROI, ugly but pragmatic.
2. Audit the 7 known flaky tests and fix them one-by-one. Each is likely a fixture-order or connection-reuse issue. Largest cleanup, but durable.
3. Bump bun test --timeout to 30s (up from our 15s). Absorbs more Blacksmith jitter. Lowest effort.
4. Quarantine the known flaky test files behind process.env.GENIE_TEST_STRICT=1. Default CI: skip them. Nightly / on-demand: run them. Gives fast merges without losing long-term signal.

Not this wish

pg-test-perf shipped the harness improvements that were the actual deliverable:

• Shared pgserve daemon (ram/template cache/lockfile reuse)
• Admin-connection reuse
• Lazy pgserve boot for non-PG tests
• macOS RAM-disk opt-in
• CI job split (unit-tests + pg-tests + umbrella)

The flake inventory is repo-wide baseline that predates the wish — confirmed by observing dev's own CI runs (e.g., run 24794680717) failing on the same kind of intermittents before this branch existed.

Assign to whoever owns the flaky test files. Tag: flaky-test, ci-health.

[namastex888]

Closing — flake mitigation landed on dev:

• 68259406 — ci(pg-tests): retry-once on bun test to mask PGlite flakes (#1335)
• 1ed61e35 — ci(unit-tests): retry-once on non-PG bun test (#1335 sibling flake class)
• ad2dbdf4 — fix(deps+doctor): pin every runtime dep + bump pgserve 1.1.10 → 1.2.0 + darwin shmem cleanup (touches the same env stack)

The 1/3626 flake-rate is now masked at the CI layer. If a deterministic root-cause fix is still wanted, please file a follow-up scoped to the specific test (resume.test.ts cross-test PG pollution, etc.).