test: drain DOM tasks before MSW shutdown

[bolichen97]

Problem / Motivation

The full Vitest suite could finish every assertion and still fail during worker teardown. happy-dom could leave an iframe or resource fetch queued until after MSW closed, producing an unhandled AbortError or a real ECONNREFUSED dial to the test origin. Node 25 workers also exposed process-global experimental Web Storage without a persistence path, emitting a warning for every test file.

Why it matters

These failures move between shards with worker timing and obscure real failures. A global rejection filter would also risk hiding genuine application defects, so the test lifecycle itself needs to be deterministic.

What changed

• register the happy-dom task drain after the MSW close hook, so Vitest's reverse teardown order drains DOM work while interception is still live
• begin timer teardown before the drain so late registrations cannot create new work
• keep shared setup compatible with suites that explicitly use jsdom and have no window.happyDOM
• remove process-level unhandledRejection suppression for AbortError and ECONNREFUSED
• replace wall-clock timer assertions with zero-delay FIFO event-loop barriers
• disable Node 25's unrelated process-global experimental Web Storage in Vitest forks; happy-dom's deterministic window storage remains in use
• structurally pin the worker and teardown contracts

Global overlap audit

All open PRs were checked for the four changed files. #4913 touches the build-plugin list in vite.config.ts; #5274 touches the coverage exclude list. This PR changes only the Vitest worker execArgv region, so neither contributor's behavior or hunk is replaced.

Tests

• six focused setup/jsdom suites: 84 passed
• npm run typecheck
• changed-test ESLint: 0 errors
• git diff --check
• clean merge-tree against current main
• full-suite reproduction after removing the rejection filter: no teardown ECONNREFUSED or process-global localStorage warning; unrelated deterministic failures found there are being fixed separately rather than hidden here

Screenshot evidence

Why no screenshot: This changes only the Vitest worker environment, shared test teardown, and deterministic test coverage; no rendered component, layout, style, motion, or interaction behavior changes.

Checklist

• One Conventional Commit
• No retry, sleep, timeout increase, warning filter, or rejection suppression
• Existing open-PR changes preserved
• No secrets or generated artifacts

[github-actions]

UX Review (Fable 5, fork) — ✅ PASS

UX-level review of a973d0c07cacd99fa25de0177e4bd5883a1074cf via the fork AI-review pipeline — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

UX-Verdict: PASS

No user-facing surface is touched — only Vitest teardown ordering, test assertions, and worker flags — so there is no rendered experience to evaluate.

[UX-REVIEWED] a973d0c

[github-actions]

Design Review (Fable 5, fork) — ✅ PASS

Design-level review of a973d0c07cacd99fa25de0177e4bd5883a1074cf via the fork AI-review pipeline — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

Root-cause fix: reorders teardown so DOM work drains under live interception, correctly replacing a process-level rejection filter that could mask real defects.

Watch

• waitUntilComplete() has no time bound, and beginTimerTeardown() now runs before the drain — so an in-flight task whose completion depends on a wrapped Node timer (e.g. an MSW handler mid-delay() when the file ends) gets its timer cancelled on the spot and the drain waits until vitest's hook-timeout. The old suppressed-rejection path was silently green there; the new failure is loud but surfaces as an opaque afterAll timeout in shared setup.ts rather than naming the leaking test.

Suggestions

• Bound the drain (race against a generous timeout) and, on expiry, throw an error reporting that DOM tasks were still pending at teardown — same fail-loud outcome, but attributable instead of a bare hook timeout.

[DESIGN-REVIEWED] a973d0c

[github-actions]

Opus 4.8 Review (fork) — ✅ no blocking findings

Reviewed a973d0c07cacd99fa25de0177e4bd5883a1074cf via the fork AI-review pipeline; updated in place on each push.

Review details

Both candidates fail falsification:

Candidate 1 (--no-experimental-webstorage breaks Node 22.0–22.3): The flag was added in Node 22.4.0, so the negation aborts the worker fork on 22.0–22.3, which the >=22 engines floor nominally permits. But Node 22 only became LTS at 22.11.0; 22.0–22.3 are superseded pre-LTS "Current" builds nobody realistically runs in 2026, and CI runs Node 24. The concrete input reduces to "if a developer were to run 22.0–22.3" — a speculative condition, exactly what the discovery pass itself rated low confidence. Below the bar.

Candidate 2 (removed unhandled-rejection guard): The discovery pass states outright it "could not construct a concrete surviving task" that escapes waitUntilComplete(). Without a concrete input (a) and observable outcome (c), it is ungrounded. Dropped.

No grounded Step 2 findings: the setup refactor, structural-pin string search (afterAll(async () => {\n await settleDOMTasksBeforeMockShutdown(window.happyDOM) and afterAll(() => server.close()) both present, drain registered after close), and the timer-teardown test rewrites all check out.

No findings.

[OPUS-REVIEWED] a973d0c

[github-actions]

GPT 5.6 Review (fork) — ✅ no blocking findings

Reviewed a973d0c07cacd99fa25de0177e4bd5883a1074cf via the fork AI-review pipeline; updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] a973d0c