[matt2e]

Fixes the two CI failures of PR #2609's run 42592827846. Neither is caused by that PR; this branch is cut from main and is independent of it.

Summary

Two test-infrastructure fixes, no product code changes:

• WebKit browser shard 9/14 hits the 15-minute job deadline with every test passing. WebKit journeys now run on eighteen shards.
• A WebKit lazy import() fails when the Vite fixture server closes an idle keep-alive connection. Fixture and preview servers no longer close idle connections on their own.

Changes

ci: run WebKit browser journeys on eighteen shards (.github/workflows/ci.yml, docs)

• Root cause: job 129078698059 "Browser journeys (webkit, 9/14)" was cancelled at the browser job's timeout-minutes: 15 with 87 of 92 tests done, all passing. The shard is structurally at the cap on main, not on one slow runner: its whole-job durations in recent runs were 15.1 min (cancelled, this run), 14.9, 14.9, 14.7, 13.7, 12.1, 10.4, 10.2 and 8.9 min, and shard 10/14 runs 13.8–14.1 min; every other shard is ≤ 12.6 min. Playwright shards by test count at file level, and the heaviest files (nested-replies 348 s, navigation-thread-history 253 s, shortcuts 194 s, right-pane-dismissal 188 s) landed together: about 25 min of summed test time on two workers plus ~1.5 min of setup.

• Fix: eighteen WebKit shards, the smallest count that leaves clear headroom under the unchanged 15-minute deadline. Chromium keeps nine; the matrix excludes shards 15–18 for it (expanded matrix: 9 Chromium + 18 WebKit = 27 jobs). The explanatory comment in ci.yml and the shard counts in docs/browser-testing.md and docs/contributing.md are updated. scripts/flake-check.mjs sizes its own lanes independently of this matrix, so its constants are unchanged.

• How the count was chosen: per-test durations from this run's browser-journeys-webkit-* artifacts (ci-report.json; shard 9's own report came from main run 42576311891, since the cancelled job uploaded none), combined with playwright test --config tests/browser/playwright.ci.config.mjs --project webkit --list --shard=K/N for each candidate. Predicted job time uses the flake-check model (105 s outside tests + 1.11 × half the shard's summed test time):

  WebKit shards	slowest shard: summed test time	predicted job time	heaviest files in it
  14 (before)	24.9 min	15.6 min (actual 15.1, cancelled; 14.9 on main)	nested-replies, navigation-thread-history, shortcuts
  16	20.3 min	13.0 min	unread, navigation-groups, agent-activity
  18	18.2 min	11.9 min	thread-unread, unread, agent-activity
  20	17.0 min	11.2 min	navigation-groups, agent-activity, initial-position

  Calibration on this run: predicted 14.7 vs actual 13.9 min (shard 10), 13.5 vs 12.4 (shard 5), 11.8 vs 10.8 (shard 3). Going from 18 to 20 shards buys 0.7 predicted minutes for two more runners, so 18 it is.

test(browser): keep fixture servers' idle keep-alive connections open (tests/browser/vite-server.mjs, tests/integration/vite-fixture.test.mjs)

• Root cause: job 129078806819 "Flake check (webkit, 5/11)": reactions.spec.mjs:8 "reaction plus opens a visible emoji-only picker…" failed repeat 3 of 10 at expect(errors.unexplained()).toEqual([]) with Unhandled Promise Rejection: TypeError: Importing a module script failed.; the other 9 repeats and all 43 other results passed. The trace (artifact flake-browser-webkit-5) has exactly one failed request: GET /src/features/messages/reaction-name.ts on the fixture server, "Connection terminated unexpectedly" after 12 ms with no send/wait/receive timings, i.e. the socket was already dead. The previous requests on that server were 4–6 s earlier. Node's http.Server closes a keep-alive connection after its default keepAliveTimeout of 5 s (measured locally: 6.0 s after the last response, with Node's timer granularity); Vite leaves the default. WebKit on Linux reused the idle connection in that instant, the lazy import died with the socket, and nothing retries a module script. tests/browser/page-errors.mjs is deliberately not touched: the app did leave a rejection unhandled, so this is not an engine report to explain away.
• Fix: keepIdleConnections sets keepAliveTimeout = 0 (Node: "disable the keep-alive timeout behavior on incoming connections") on every fixture dev server, in the fixture plugin's configureServer next to bindAssignedPort so it survives server.restart(), and on every preview server in configurePreviewServer. headersTimeout stays at its 60 s default. Vite destroys open sockets when a server closes, so nothing leaks. A new integration test asserts the setting on a created server, after a restart, and on a preview server; it fails against the unmodified module.

Verification

• Matrix: expanded ci.yml's browser matrix with the yaml package: 27 jobs, 9 Chromium (1/9…9/9), 18 WebKit (1/18…18/18). playwright test --list --shard=9/14 with and without --no-deps gives the same 92 WebKit tests, so the local listings above match what CI shards.
• node --test tests/integration/vite-fixture.test.mjs: 9/9 pass; the new test fails (1/1) with tests/browser/vite-server.mjs reverted to main.
• biome check on the two changed .mjs files: clean.
• tests/browser/reactions.spec.mjs in WebKit through the CI config against the fixed server: 4/4 pass (one run, one worker).
• Mechanism, Node level: a plain http.Server with the default closes an idle keep-alive socket 6.0 s after the last response; with keepAliveTimeout = 0 it is still open after 9 s and the client agent reuses it.
• Not reproduced in a browser: a probe that loads a page from a fixture server, fetches reaction-name.ts, idles, then import()s it never failed in local WebKit (macOS) before the change, including a 2.5 ms sweep from 5.975 to 6.0325 s around the measured close and coarse gaps of 3–8 s. macOS WebKit uses CFNetwork, which retransmits an idempotent request on a stale connection; the CI failure was Linux WebKit (libsoup), which the probe cannot exercise here. The fix therefore rests on the trace evidence and the Node-level mechanism, not on a browser fail-then-pass. After the change the same probe reports keepAliveTimeout 0 and passes at 6 s and 8 s idle.
• Not run: the full browser suite or a WebKit run on all 18 shards; CI on this PR provides the actual shard durations.

🤖 Generated with Claude Code

[wesbillman]

🤖 Star Lord’s automated source review via Wes’s account, with independent Rocket, Mantis and Groot coverage.

Reviewed head 7ab8b403bb43e61ca043e59ffcb572f1ed6e4f97 against base/merge-base 58116b826171d190a7376fc57b0509c32c507dde. No blocking source defects found in the shard matrix, fixture keep-alive lifecycle or regression assertions.

Optional documentation correction: docs/browser-testing.md:142 still says “twenty-three Linux selections.” Update it to “twenty-seven” to match the new 9 Chromium + 18 WebKit matrix and the surrounding documentation.

Hosted evidence inspected: run 42597212482 passed all 27 browser shards. The slowest WebKit job was 1/18, at 10m37s including setup/teardown; its timing artifact reports 66/66 passes, 9m26s test wall time and 17m41s summed test execution. The slowest file was appearance.spec.mjs at 207s summed execution. These results support headroom in this run, not a guarantee across runners.

Overall CI remains failing: Rust and tool integration stopped at dependency policy on libsqlite3-sys / RUSTSEC-2026-0459, before the Node integration tests. This PR leaves Cargo.lock and deny.toml unchanged. The new integration test checks the timeout setting across creation/restart/preview; it does not reproduce the Linux WebKit race. The PR’s local tests, original trace analysis and before-change timing claims remain author-reported. This review executed no code, tests, probes or apps.

[kazemek21]

Kyle reviewed this PR and the Nexus review-tracker notes and approved with no comments.

— Posted by Codex (gpt-6.1-sol) on behalf of kylez