[matt2e]

Summary

Fixes three test flakes from the 29–30 Oct CI flake sweep:

• Typing lost in a dialog that opens while a menu is still closing.
• A history-loading journey that never left the bottom of the timeline.
• A heavy-tailed signing loop in unread.test.ts.

No product code changes. This is one of six independent PRs split from the flakey-tests branch. Each is cut from main and can merge in any order.

Dialog typing lost after a menu (navigation-groups.spec.mjs:865, solo-chats.spec.mjs:243)

• CI: main 41003205039 (job 123892389615, WebKit 10/12), where :917 got "" for "Retained launch". [5/5] Add protection to the Goose harness plugin #839 run 40964340233 (job 123765272909, WebKit 8/12), where "Create and move" stayed disabled with an empty name.
• Cause: fill() focuses the field, then sends insertText as a separate call. If focus moves in between, WebKit drops the text without an error. Both traces show fill() returning with the value still "" on an input that did not remount. At that point the dialog's initial focus (layout effect, microtask, then rAF) is still pending, and NewWorkspaceDialog mounts while both Solo menus are still exiting. The traces do not show what moved focus in each run.
• Change:
  • navigation-groups waits for toBeFocused() before fill() and checks toHaveValue() after, as navigation-group-icons:238–242 and the sibling test at :720 already do. The :917 check now tests only that the draft survives the failed refresh.
  • The three Solo New workspace flows (solo-chats :105 and :249, solo-rename :348) first wait for page.locator('[role="menu"]') to reach count 0, then toBeFocused(), fill() and toHaveValue(). The locator is CSS because the modal sets aria-hidden on the exiting menus, which hides them from role queries.
• Verified: a temporary control (not committed) set initialFocus={false} on NewWorkspaceDialog and CreateSidebarSection. With it, every new barrier fails at toBeFocused() in both engines. Without it, all four tests pass in Chromium and WebKit. The WebKit race itself was not reproduced locally.

history-loading.spec.mjs:593 "a live message after a failed older load is new"

• CI: 40980132056 (job 123818199988, WebKit 2/12) on feat(identity): the first-run Welcome flow replaces the key-setup gate (FOUNDATION: App.tsx) #1820, which touches only identity code and App.tsx. It timed out at :610 waiting for "1 new message".
• Cause: in the trace, page.mouse.wheel(0, -100000) moved the reader only 41 px. loadNearTop starts an older page at any scrollTop below max(3000, 4 × clientHeight), so the quota-failed load and its "rate-limited" alert still appeared. updateJumpToLatest treats anything within 80 px as the bottom, though, so the latest-messages button never appeared. The test had taken the alert as proof that the reader had left the bottom. Why WebKit stopped at 41 px is not determined.
• Change: after arming the quota, the test calls upper() from timeline.mjs instead of the hover and raw wheel. upper() requires each bounded gesture to make progress until the reader is more than 400 px above the bottom. The alert wait, settle, publish and assertion are unchanged. The test at :126 already establishes that the extra scrolling does not retry a quota error.
• Verified: a temporary control (not committed) clamped the first upward wheel gesture to 40 px, as timeline-setup.spec.mjs:17–25 does. The old test then stopped 40 px from the bottom and failed at the latest() assertion in both engines. The new test ended 650 px above the bottom and passed in both. Without the control, it passes in both.

unread.test.ts "inbox describes a row by its newest message within the same second"

• CI: 41010078970 (job 123913747149, JavaScript 1/3) on feat(github): unfurl PR and issue links inline with their title #1693, which touches no unread files. It failed with "Test timed out in 5000ms", reported after 16,678 ms.
• Cause: pair() signed done once and re-signed only working until working.id > done.id. With done.id fixed, the chance of needing more than n tries is 1/(n+1), a heavy tail, and the test makes two pairs. The loop is synchronous, so Vitest reports the timeout only once it ends.
• Change: every try re-signs both events, with ms tags 100+i and 900+i. Both keep created_at 20, and working's ms stays below done's, so Done still follows Working by its ms offset alone. Each try now succeeds with probability 1/2.
• Verified: the file passes 150/150. A scratch script hashed the test's event shape with getEventHash. Over 1,000 pairs (twice) and 10,000 pairs, the old loop's maximum grew with the sample (818, 18,777 and 242,109 tries). The new loop's maximum stayed at 11–15, with no pair over 20 tries.

Overlaps

• fix(messages): preserve held readers through equal-height resizes #1963 and fix(messages): preserve held readers outside the mounted range #1956 edit ChannelTimeline.tsx and history-loading.spec.mjs. They own the history-loading :298 and :334 fixes, which were dropped from this branch. Neither touches updateJumpToLatest, loadNearTop or the :593 test.
• fix: restore row focus when escape beats menu focus #1942 edits ui/finalFocus.ts, but only for an Escape race.

Browser coverage

No browser cases were added or removed. Four journeys gain focus and menu-unmount barriers before typing. history-loading:593 now reaches its reading position through the shared upper() helper instead of one raw wheel.

Verification scope

tsc --noEmit and biome check ran on the five changed files. Playwright ran only the five touched tests, one single-worker process at a time: Chromium, then WebKit (5/5 each). The controls ran the same way and were reverted.

Remaining gaps

• Hosted CI has not run yet.
• A green run does not show that these flakes are gone.

🤖 Generated with Claude Code

[kazemek21]

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

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

[wesbillman]

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

No actionable findings in the five-file test-only change. The focus barriers retain draft/recovery assertions, the shared upper() helper establishes an above-bottom reading position before the live-message check, and re-signing both events preserves the test’s opposing ID/millisecond ordering. No production code or browser cases were removed.

Reviewed head 037888efe392dfd09a669c8652ea8ec324727a18 against base 36a283b2386473130430890faa3b06c6e45d4a6b, using verified Git blobs. The public description and commit metadata were inspected; there are no attached images and no new concrete publication-surface finding.

Hosted CI run 41051678204 succeeded on merge 8ab1d45dcdd79400d5d535327114b21604317f78, whose tree matches the reviewed head. The changed history case passed ten advisory repetitions in each browser engine; the full unread file passed 150 tests. Counts, elapsed times and slowest-case evidence were inspected, but no equivalent base timing comparison was obtained.

Source-only review: no local tests, builds, app launches or PR-code execution. The author’s temporary failure controls/statistical experiments were not independently reproduced; the original WebKit input cause, elimination of flakes and before/after cost improvement remain unproven. This COMMENT is feedback, not approval or merge authorization.