test: stop codex app-server runtime tests flaking under parallel load

[danshapiro]

Problem

test/unit/server/coding-cli/codex-app-server/runtime.test.ts flakes in the broad server-config run (it passes in isolation). The hand-off blamed the per-test 3000 ms wall-clock timeout, but a contention storm (60 CPU-busy loops + 16 concurrent copies of the file) showed that's only a small slice:

Failure mode	Share
ensureReady() rejected — did not finish initialize within Nms / connect ECONNREFUSED	dominant (~32/34)
Test timed out in 3000ms (wall-clock)	~2/34

Root cause: these tests spawn real node sidecars (the fake-app-server fixture) and require them to spawn + bind a WebSocket + answer initialize. The server-config suite runs with fileParallelism + maxConcurrency (vitest.server.config.ts), so freshly-spawned sidecars are CPU-starved and can't meet the tests' tight per-attempt budgets (startupAttemptTimeoutMs of 200/500 ms), so ensureReady rejects at the await. It's a family, not one test: retries startup when the preallocated loopback port is lost (200 ms), keeps the same sidecar when wrapper identity is transiently incomplete (500 ms), uses the startup attempt timeout to tear down an initialize hang (500 ms + a Date.now()-start < 1500 assertion), and both tears down the owned process group before retry… tests (500 ms).

Production is unaffected — DEFAULT_STARTUP_ATTEMPT_TIMEOUT_MS = 3000; only the tests shrank it for speed. This is test-config fragility, not a product bug, so the change is test-only.

Fix

• Real-success startup tests → contention-tolerant per-attempt budget (REAL_STARTUP_ATTEMPT_TIMEOUT_MS = 5000, above production's own 3000 default) and drop the redundant }, 3_000) wall-clock overrides so they inherit the 30 s global.
• Initialize-hang test → keep its tight budget (it verifies the per-attempt timeout fires), but turn the absolute < 1500 ms into a boundedness check tied to the configured budget (limit * timeout + slack).
• Wait helpers (waitForPidFile / waitForProcessExit / waitForMetadataRecord) → raise the poll deadline (WAIT_HELPER_TIMEOUT_MS = 15000); they observe second-order fixture spawns (e.g. a native child) that lag under load.
• writes ownership metadata immediately after spawn… → wait for the identity-complete metadata write via a predicate instead of racing the runtime's first (empty-identity) write.

Verification

• In isolation: 53/53.
• Under the storm (16 concurrent copies × 53 tests, 60-process CPU load, load avg ~135 — heavier than the pre-fix run): 16/16 copies fully green, 0 failures, vs 5 distinct tests flaking pre-fix.
• typecheck:server and typecheck:client clean; targeted tsc on the test file clean.

No production code touched; change is confined to one test file.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.