test: resume.test.ts 'manual resume: suspended agent with sessionId is resumed' fails ONLY in full suite (cross-test PG pollution)

[namastex888]

Summary

src/__tests__/resume.test.ts:109 test case resume > manual resume: suspended agent with sessionId is resumed passes reliably in isolation (10/10 runs, 3.6s) but fails consistently when invoked via full bun run check (3/3 attempts on 2026-04-21 at 10:40-11:05 UTC on dev HEAD 39bf0290).

Classic cross-test PG state pollution — not a load-dependent oscillating flake like the baseline-flake family (listen-bomb, spill-journal, SpawnTargetPicker).

Repro

cd /home/genie/workspace/agents/genie-configure/repos/genie
git checkout origin/dev  # 39bf0290 (v4.260421.12)

# Isolated — PASSES
bun test src/__tests__/resume.test.ts
# → 10 pass / 0 fail in 3.6s

# Full suite — FAILS consistently
bun run check
# → 3415 pass / 1 fail in ~128s
# → fail: "(fail) resume > manual resume: suspended agent with sessionId is resumed"
# → underlying error: postgres handler at node_modules/postgres/src/index.js:331

Impact

• Blocks all local pushes via .husky/pre-push (full bun run check).
• First hit during attempted ship of genie brain <subcommand> exits 0 when brain is not installed — silent data loss for write subcommands #1151 fix (branch fix/1151-brain-subcommand-exit-nonzero, unrelated 3-line edit in brain.ts).
• Same blocker surface as bug(test): prependEnvVars import fails under concurrent test runs, blocking all local pushes #1223 (prependEnvVars) was before — now a 4th flake family in the pre-push-blocker bestiary.

Root cause hypotheses

1. Shared pgserve port 20900 not reset between test files — src/__tests__/ and scripts/reconcile-orphans.test.ts both use pgserve --ram on port 20900. State bleed.
2. FK cascade from sibling test — another test inserts an agent row that interferes with resume.test's suspended-agent fixture on cleanup.
3. Test runner ordering — isolated run picks different order; full run exposes dependency on a prior test's rollback.

Related

• brain/memory/reference_baseline_flakes_2026_04_20.md — the existing 3-flake family
• brain/memory/reference_baseline_flakes_blocker_2026_04_21.md — my own blocker doc during this session (but this is a DIFFERENT flake than listen-bomb/spill-drain)
• bug(test): prependEnvVars import fails under concurrent test runs, blocking all local pushes #1223 (prependEnvVars) — closed 2026-04-21, same "full-suite-only" fail signature

Suggested next step

bun test src/__tests__/resume.test.ts --preload=<setup-that-forces-fresh-pgserve> to isolate. If isolates, fix is test harness-level (per-file pgserve reset, or serialize PG-touching test files).

Reported by the bugless-genie orchestrator during autonomous overnight work, 2026-04-21.

[namastex888]

overnight-genie-issue-loop status: BLOCKED after 2 cycles. Engineer thrashed re-running the flaky test itself rather than designing a fix; repeated timeout/kill cycles with no branch pushed. Per wish two-strikes rule, requeued for fresh engineer or human design pass. Worktree was cleaned. PG-isolation root-cause noted in #1335 (baseline flake rate 1/3626 surfaced by pg-test-perf wish).

[namastex888]

Effectively resolved by fd74f49 "fix(emit): latch shuttingDown across setupTestDatabase — close leaked-poller gate".

That commit fixed the exact singleton-leak shape my draft wish (fix-resume-test-cross-suite-pollution) had identified as Hypothesis 2: the emit.ts shuttingDown latch leaking across setupTestDatabase calls, leaving background pollers (audit.ts:193 setInterval, runtime-events.ts:343 subscribe+poll) firing emitEvent → admitToQueue → ensureFlusher against a soon-to-be-dropped DB connection, propagating CONNECTION_ENDED into the next test file's first await.

Fix landed:

• shutdownEmitter() keeps shuttingDown=true latched on return; resumeEmitter() is the explicit re-open
• ensureFlusher gates on shuttingDown; admitToQueue early-returns when shuttingDown
• doFlush() catch returns early instead of requeueing when shuttingDown (prevents queue repopulation after shutdownEmitter clears it)
• test-db.ts calls resumeEmitter() after resetConnection and at end of cleanup

Verified live: bun test src/__tests__/resume.test.ts → 11 pass / 0 fail. The wish I drafted (.genie/wishes/fix-resume-test-cross-suite-pollution/) is now stale and can be archived/discarded — the underlying fix is upstream of any work my dispatch produced.

Closing.