Flaky server-unit 404s under full-suite runs: routes intermittently unmounted

[bokelley]

Two distinct server unit tests have failed precommit full-suite runs today with the same signature — an HTTP route test that expects a real status gets 404, and the same file passes in isolation:

• server/tests/unit/registry-brand-setup-route.test.ts > rejects unsafe logo URL — expected 400, got 404 (passes 22/22 alone)
• server/tests/unit/slack-webhook-signature-gate.test.ts > answers only valid signed URL verification challenges — expected 200, got 404 (passes 13/13 alone)

Both under vitest run of the full server-unit suite (~434 files / 6.1k tests), different test each occurrence, ~2 of 4 full-suite runs affected on one machine today. The 404 shape suggests order-dependent route mounting or shared app/server state across test files (a race in app construction or a file that tears down/replaces the router), not assertion drift.

Impact: intermittent precommit/CI failures on unrelated diffs (hit repeatedly while landing PR #6549's docs-only changes).

Repro attempt: npx vitest run --dir server/tests/unit --pool=threads in a loop; failures are probabilistic.

No workaround applied — flagging for triage per the file-issues-for-workarounds policy. Removal criterion: 20 consecutive full-suite runs with no 404-flake.

[bokelley]

Triage

Classification: Bug
Bucket: infra / agents (server test tooling)
Status: ready-to-implement

Code analysis:

• server/vitest.config.ts sets pool: 'forks' + fileParallelism: false — isolation intent is right.
• scripts/precommit-server-unit.cjs always passes --config server/vitest.config.ts; precommit does not override the pool.
• package.json test:registry runs vitest run server/tests without --config — a secondary risk if used in CI, as it gets vitest's default pool rather than the forks config.

My take: Two layered issues. The reporter's --pool=threads repro is accurate: threads share a Node.js VM, so vi.mock() mutations in one file bleed into another's router construction and drop routes (404s). The actual precommit failures (which do use the forks config) are likely from vitest 4.x's forks pool reusing fork processes between sequential files without fully resetting the module registry — isolate: true is not explicitly set and may not be the v4 default.

Ready to implement:

• server/vitest.config.ts: add poolOptions: { forks: { isolate: true } } to guarantee a fresh module registry per file even under sequential fork reuse.
• package.json: add --config server/vitest.config.ts to the test:registry script so it inherits the forks isolation settings.
• Both failing test files (registry-brand-setup-route, slack-webhook-signature-gate) are correctly written — no changes needed to the tests themselves.
• No changeset needed (server test infra, not protocol surface).
• Removal criterion from issue: 20 consecutive clean npm run test:server-unit runs.

---

Triaged by Claude Code. Session: https://claude.ai/code/session_014zcC6epiNYdD57xNpTc4hD

---

Generated by Claude Code