[matt2e]

Fixes a WebKit flake in thread-composer-focus.spec.mjs:168, "a link to a specific reply reveals it without focusing the composer". It failed in WebKit shard 10/14 on two unrelated branches within 36 h: run 41494357757, job 125445893123, and run 41529537399, job 125566410581. The second one passed when the same SHA was retried. Both runs failed at :202 with toBeInViewport on the thread row reporting "element(s) not found".

Cause

The click on the link got lost, so the thread never opened. The product behaves correctly: a real press and release on the link opens the thread.

• The test clicked the newly appended message's link before its row had settled. The row first mounts unmeasured (visibility: hidden). When the byline resolves, the link label changes from "Alpha" to "Alice Fixture in #Alpha", and the timeline follows both changes. In both traces, Playwright's first click attempt failed with "element is not stable" or "element is not visible".
• Each retry calls scrollIntoView with forced alignment, which really scrolls the timeline (2192 → 2214 px). Under CI load, the resulting scroll event was handled after Playwright's press but before its release.
• Virtua treats that event as a user scroll and sets pointer-events: none on the list for 150 ms. The release then lands on the scroller section, so the link's onClick never runs and the URL never gains messageId.

Change

Before any pointer input, the test now finds the link by role using its resolved name and waits for it to be visible, then waits for virtuaIdle. buzz-links.spec.mjs already uses this pattern for the same problem. Product code is unchanged.

Verification (local macOS, pinned Playwright, --no-deps)

• Fail-before: I built a deterministic repro and did not commit it. An init script delays the timeline row's ResizeObserver callbacks by 450 ms. It also holds the forced-alignment scroll event until the press's mousedown, which matches the 85–114 ms click step seen in CI. With this injection, the unchanged test failed 10/10 on WebKit with the CI signature: the release lands on the section, the list has pointer-events: none, and there is no messageId.
• Pass-after: With the same injection, the changed test passed 10/10, and no forced-scroll retry happened.
• Control: Waiting for the visible, named link without virtuaIdle also passed 4/4 under the injection. virtuaIdle is kept to cover the re-measure after the name change and Virtua's hold, which intercepted the attempt-1 press in one CI trace.
• Repeat runs: The changed test passed 30/30 on WebKit with --repeat-each=30, 24/24 with 6 workers, and 12/12 on Chromium with 4 workers. All 7 cases in the file pass on both engines.

Browser cases added: 0. Removed: 0.

Deferred checks

• Hosted CI has not run on this change yet.
• A green run does not prove the flake is gone.

🤖 Generated with Claude Code