test: silence the 8,000-line CI test log — settle the router inside act, stub jsdom's no-ops, quiet the hub's counter feed in tests

[karngyan]

The noise

The v0.1.0 release run's "Lint and test" step printed ~8,500 log lines, ~95% noise. Measured profile (run 31333105025):

Class	Lines	Source
React act(...) warnings	736 warnings × 7-line block ≈ 5,150	Transitioner (314), MatchImpl (145), MatchInnerImpl (143), MatchesInner (104) — all TanStack Router internals — plus Root (16), UpdateNotice (8) and a handful of others
jsdom Not implemented: Window's scrollTo()	65	focus mode / Radix scrolling in jsdom
jsdom canvas getContext()	8	the Devices QR canvas in jsdom
relay {"evt":"channel_closed",...} JSON	76	the hub's deliberate Workers Logs counter feed, leaking through relay/test/
jsdom navigation to another Document	3	link-click navigation in jsdom

None of it showed locally because vitest's local default reporter hides console output from passing tests; CI's reporter prints it. pnpm vitest run --reporter=verbose reproduces the CI view exactly (735 warnings locally vs CI's 736), which is what the before/after numbers below are measured with.

The fixes, at their seams

act warnings — fixed in the render helpers, not per-test. Every router-mounting helper did await router.load() and then a bare render(<RouterProvider .../>). Mounting RouterProvider re-runs router.load() from Transitioner's mount effect, and that call's continuations update the router stores a microtask after RTL's synchronous act has exited — one warning per mounted Match, every mount. The helpers now render inside await act(async () => ...) so the post-mount load settles inside act:

• web/src/testing/render.tsx (renderWithRouter — covers app-shell, session-table, page-header, devices, remote suites)
• web/src/router.test.tsx (renderAt, renderFleet, renderPicker, renderBare, one inline render)
• web/src/routes/sessions.test.tsx (mountSessions, plus one attached that starts a navigation)
• web/src/routes/pair.test.tsx (renderPair, renderRelayPair)
• web/src/components/app-shell.test.tsx, web/src/components/update-notice.test.tsx: their helpers emitted welcome frames through the synchronous FakeSocket outside act; wrapped.

No global suppression: IS_REACT_ACT_ENVIRONMENT untouched, no console filtering of React output — an act warning from an app component still prints.

jsdom no-ops — stubbed in web/src/testing/test-setup.ts, in the same shape as the existing scrollIntoView stub: window.scrollTo becomes a no-op, HTMLCanvasElement.prototype.getContext returns the null jsdom already returned after logging (the Devices no-canvas fallback test still pins that path).

relay counter line — silenced in tests only. The channel_closed line is production observability (relay/src/hub.ts, retireClient) and stays byte-identical. A new relay/test/setup.ts drops exactly that prefix at the printing edge; everything else the Worker logs still prints. The counters test (test/hub.test.ts, "logs the counters") passes unchanged — its vi.spyOn(console, 'log') wraps the filter and records the call either way.

Before / after

Measured with the reporter mode that matches CI's output (--reporter=verbose, prints passing tests' console output):

Surface	Before	After
web act warnings	735 (CI: 736)	0
web jsdom noise lines	76	3 (see below)
web total (verbose)	8,661 lines	1,010 lines (≈998 of them the per-test ✓ listing)
web total (default reporter, pnpm vitest run 2>&1 | wc -l)	85	12
relay channel_closed lines	76	0
relay total (verbose)	201 lines	101 lines

Tests: web 56 files / 998 tests, relay 4 files / 90 tests — all passing before and after, none skipped. pnpm run lint (tsc) clean in both; styles.build.test.ts passes (all new comments live in *.test.tsx / src/testing/ / relay/test/, outside the Tailwind scan perimeter).

Deliberately left noisy

• 3 × jsdom Not implemented: navigation to another Document — the only seam is jsdom's link-click navigation itself, and suppressing it (global preventDefault, anchor click stubs) would change what the anchor-click tests prove: the router-Link-vs-plain-<a href> distinction in testing/render.tsx depends on jsdom not navigating.
• 1 × workerd's own disconnected: read end of pipe was aborted in the relay run — runtime-internal, not ours to filter.

🤖 Generated with Claude Code

[karngyan]

Confirmed against this PR's own CI run (31334243256), same runner class as the v0.1.0 release run the profile was measured on:

"Lint and test" step	v0.1.0 release run	this branch
step log lines	8,014	240
act(...) warnings	736	0
channel_closed lines	76	0
scrollTo / getContext lines	65 / 8	0 / 0
web tests	998 passed	998 passed
relay tests	90 passed	90 passed

One caveat worth naming: the first attempt of this run failed on relay/test/hub.test.ts > the handshake deadline > re-arms while unseen clients remain, then reaps them too — the alarm had not re-armed because both clients were already past their 50 ms deadlines when the first alarm ran, i.e. the runner stalled somewhere in a 25 ms window. That is the pre-existing real-time-sensitive reap test (TIMEOUT_MS = 50, sleep(40)/sleep(25) margins), the same CI-only failure class the comment atop relay/vitest.config.ts documents; nothing in this branch touches alarms or timing, and the untouched re-run passed. Left alone here as out of scope, but it will flake again on a slow enough runner.