Tests flake under parallel load: hard 5s timeouts lose races against CPU contention #53

Pre-existing flake on fork/main — not introduced by any open PR.

Symptom

When several full npm test runs execute concurrently, unrelated tests fail at almost exactly 5000 ms:

• withAppServer falls back to a direct app-server on broker/busy after a successful handshake (#32) — 5003 ms
• broker CLI: exits with code 1 when app-server child dies (ACK-based) — 5006 ms

The ~5000 ms figure and the fact that which test fails varies per run point at a hard 5-second timeout being exceeded under CPU contention, not at a logic defect.

How it was isolated

Ran the same tests in three states on macOS (Node v25.9.0, codex-cli 0.146.0), using separate git worktrees:

Reproduction of the load condition: three full npm test runs started simultaneously, one per worktree. All three failed — including the untouched baseline — each on a different test, each at ~5000 ms.

A narrower load probe (6 parallel runs of a single test) did not reproduce it; the failure needs the full suite running concurrently, i.e. many app-servers/brokers competing at once.

Conclusion

The flake belongs to the test suite itself, not to any PR under review. Serialized runs are green everywhere.

Possible directions

• Make the 5s timeouts adaptive, or raise them when the suite detects contention.
• Gate the broker/app-server-heavy tests behind a concurrency limiter so they do not race each other.
• Alternatively, mark the suite as serial-only and document that concurrent npm test runs are unsupported.

Filed as a follow-up; no change proposed here.