[matt2e]

Summary

mentions.spec.mjs now builds tests/fixtures/mentions.html once in beforeAll and serves the bundle with build and preview from vite-server.mjs. Before this change, each of its four tests started its own Vite dev server. This is the same fix #2566 made for voice-notes.spec.mjs. The change is test-only: no assertions change, the navigation timeout is unchanged, and no browser cases are added or removed.

Problem

Flake check (chromium, 5/9) on #2396 failed mentions.spec.mjs:34 "actual composer selects namesakes by exact key…" in repeat 3 with TimeoutError: page.goto: Timeout 15000ms exceeded at :57. #2396 changes only a comment in this spec.

• https://github.com/block/buzz-app/actions/runs/42155727571/job/127689552962 (merge of 7d3ed5a)

The same cold-load timeout hit dev-served fixture pages on unrelated branches between 2026-11-05 20:10Z and 2026-11-06 08:10Z. All but one of those failures were in Flake check lanes, which repeat each related test 10 times:

• mentions.html via mention-edit.spec.mjs:3 (peon/profile-disk-cache): https://github.com/block/buzz-app/actions/runs/42134373233/job/127621615224
• mentions.html via completion-work, composer-recipients and mention-rules (WebKit, morganm/verified-only-people-search): https://github.com/block/buzz-app/actions/runs/42135492502/job/127624304447
• mentions.html via channel-completion.spec.mjs (am-115-attention): https://github.com/block/buzz-app/actions/runs/42146757907/job/127660281707
• mentions.html via typeahead.spec.mjs:28 (broken-8): https://github.com/block/buzz-app/actions/runs/42169660635/job/127734602256
• conversation.html, profiles.html and / on tho/plugin-member-picker, nexus/add-native-claude-execution-on-blo and peon/r3-header-members

Cause

The failing trace shows the page already rendered (composer and toolbar in the snapshot), but its load event had not fired. The per-test dev server served the fixture as 1,590 modules. Tailwind's globals.css took 9.3 s to transform, and the module-by-module waterfall ran until the 15 s budget ran out. Meanwhile, the shard's other worker was running huddles.spec.mjs. The first repeat in the same worker passed, but took 26.1 s.

Fix

• beforeAll builds the fixture page into a temp directory, and preview serves it on an OS-assigned port. afterAll closes the server and removes the directory.
• As in voice-notes.spec.mjs, the build defines import.meta.env.DEV and process.env.NODE_ENV as development, so StrictMode's checks still apply.
• The dev server's watch.ignored option is dropped, because a built page has no HMR. Tests 1–3 lose their try/finally blocks, which only closed the server. Test 4 keeps its finally for context.close(). Most of the diff is that re-indentation; git diff -w shows +57/−107.
• Flake-check selection is unchanged: moduleGraph still links the spec to tests/fixtures/mentions.html.

Validation

All runs were local on macOS (16 cores), Playwright CI config. These are not hosted CI measurements.

Fail-before / pass-after. A temporary vite-server.mjs probe (not committed) held the globals.css transform for 16 s in both the dev server and the build, the method #1971 used:

spec	result
main	:34 ✘ page.goto: Timeout 15000ms exceeded (16.9 s; CI's failure took 16.7 s)
this branch	:59 (the same test) ✓ 3.5 s; the 16 s lands in beforeAll

Cold navigation to mentions.html?test-controls, Chromium, standalone probe:

	dev server per load	build + preview
requests	1,589	10
idle	3.0–4.5 s	0.25–0.30 s (build 2.8 s, once)
24 busy loops	7.0–9.3 s	0.28–0.29 s (build 5.4 s, once)

Full file:

• Chromium: 4/4. WebKit: 4/4.
• Chromium --repeat-each=10 --workers=2: main 40/40 in 1.9 min (321 s user CPU); this branch 40/40 in 54.1 s (150 s user CPU).
• With 24 busy loops, --repeat-each=5 --workers=2: Chromium 20/20 (43.7 s), WebKit 20/20 (48.7 s).
• tests/integration/vite-fixture.test.mjs and tests/integration/flake-check.test.mjs: 33/33.

Browser coverage

No browser cases added or removed and no assertions changed. Only how the fixture page is served changes.

Remaining gaps

• Other specs that load mentions.html, conversation.html or profiles.html from their own dev servers can hit the same cold-load timeout (runs above). They are not changed here.
• Hosted CI has not run on this branch yet. A green run does not show the flake is gone.

🤖 Generated with Claude Code

[wesbillman]

🤖 Star Lord’s automated source review through Wes’s account, integrating independent Rocket, Mantis and Groot review lanes.

No actionable defects found in the fixture-serving change. The four browser cases, assertions, URL modes and navigation budget are retained. The change reuses the existing build/preview helpers and voice-notes lifecycle pattern, while browser contexts and mutable fixture state remain isolated.

Reviewed head 19ed67c0e544a3b6fc3481fad834af64b3d7fba0, target base 3bcf1c4bb1111d924a84d9b710c1da2db80ccf0b, merge-base fd21dba4af2edf63d7b038e3f0f6bdf3703b5df7.

Hosted evidence is now available beyond the PR description’s local measurements: CI run 42217003622 succeeded on synthetic merge d306214cf471dcc891d3faabaae6c5bddc7a4b4d. Its flake-check timing artifacts record all four mentions cases repeated ten times with two workers:

Engine	Passed	Elapsed wall time	Summed test execution	Slowest case
Chromium	40/40	142.6 s	198.5 s	11.0 s
WebKit	40/40	166.9 s	247.6 s	17.6 s

Both reports have zero failures, skips or flaky results. The slowest case in each is the namesake-selection/publication journey; mentions.spec.mjs is the only selected file. The linked earlier failure log corroborates the 15-second navigation timeout, but its different test selection is not a controlled before/after cost comparison.

Source review and existing CI artifact inspection only: we did not execute the PR or reproduce its local benchmark/fault-injection measurements. These successful repetitions do not prove that every intermittent timeout is eliminated.