[stamp-build]

^{Lagoon from raft opened this pull request through Stamp
View in Stamp →}

---

The emulator test thread-newer-auto-pagination failed at random in the "Android first-character baseline (emulator)" job with Timed out waiting for stable retained message AUTO-MSG-0008. This happened three times today: runs 39939224593 and 39951797674, plus one earlier run. A rerun always passed. This PR fixes the test harness. The product code is not changed.

Root cause

The test reads message positions through UiAutomation, which answers from its accessibility cache. Compose clears that cache only through throttled content-change events. Right after the first toward-newer swipe, cached rows can still report where they were before the swipe. On a local emulator, a fresh read put AUTO-MSG-0011..0013 at y=341/554/946, while the cache still had them at y=1629/2021/2413, overlapping the rows that had scrolled in.

The test picks its "retained anchor" from those reads. In local runs of the unchanged test, the anchor it picked jumped between AUTO-MSG-0008, 0009, 0010 and 0014 for the same scroll. It picked 0008 once locally, the same row as in the CI failures. 0008 is the focused row that the swipe had already scrolled away. The test then compared stale bounds with stale bounds. It passed while the cache stayed stale, and timed out when the page append refreshed the cache and 0008 was no longer in the window. The product kept its anchor: the restore logs show the newer-page settlement restoring the pointer-settled anchor every time.

This is not a scroll/settle race and not a timing wait that depends on emulator speed. It also does not come from #3578 or #3555. The fixture's realtime catch-up is a no-op, and the failure is in the harness's reads. The CI history I could pull (test.yml runs from 10-22 18:00 through today) shows this test passing in all but two of the roughly 27 full runs that included it. #3192 already reported it as flaky on 10-20, before either PR.

Fix

• Message markers are collected in one walk of the window. The walk calls AccessibilityNodeInfo.refresh() on every node and drops nodes that no longer exist, so bounds are never served from the stale cache.
• The retained anchor is the topmost marker that sits fully inside the list viewport. A clipped marker has its top pinned to the viewport edge, which hides displacement.
• The Retry and Back-to-bottom nodes are refreshed before their bounds are used.
• The two stable-anchor waits report the latest sample, the visible markers and the accessibility tree when they time out.

The waits stay bounded monotonic polls. No timeout was raised and no retry was added. There is no deterministic hook here because the property under test is the on-screen position, and only the accessibility tree reports it to the test process.

Verification

Local runs, on an API 35 emulator limited to 2 host cores with one busy loop on those cores:

• Unchanged test: 15 of 15 passed, but the anchor it read jumped between rows, as described above.
• This change: 15 of 15 passed on the old base. After the rebase onto current main, which includes fix(thread): a failed send no longer freezes the read marker, new-messages pill and tail follow #3605 and fix(realtime): Android realtime back on socket.io-client-java over OkHttp #3602, see the follow-up comment. The anchor was always a fully visible row at its live position.

Full CI (test.yml, scope=full) on head 2aaccf0 has been dispatched three times: on this branch and on two extra refs with the same commit, because the workflow cancels earlier runs on the same ref. Results will follow in a comment.

No changeset, because this change only touches a test.

🤖 Generated with Claude Code

[stamp-build]

Verification results for head 2aaccf0, rebased onto main with #3605 and #3602.

CI, full test.yml: thread-newer-auto-pagination passed in every run where the emulator suite ran.

• 6 of 6 passed with this change. Runs 39968433652, 39968741589 and 39968750143 were on the pre-rebase commit a9f2915. Runs 39982731725, 39982737911 and 39990199373 are on 2aaccf0. In run 39990199373 the anchor was AUTO-MSG-0011 at its live position, fully inside the list (top 456).
• Run 39982721749 is not counted. Its emulator jobs failed before the emulator booted ("Error on ZipFile unknown archive" while preparing the system image), so the suite never started. Stamp would not rerun it because a newer run existed on the branch, so I dispatched 39990199373 instead.
• Before this change: of the roughly 27 full runs since 10-22 18:00 that ran the test, 2 failed with the AUTO-MSG-0008 timeout (39939224593 and 39951797674). feat(cmp): big switch — whole app UI from Kuikly to Compose Multiplatform #3192 also reported it flaky (2/3) on 10-20.

Local, API 35 emulator on 2 pinned cores:

• Unchanged test: 15 of 15 passed, but the anchor it read jumped between rows 8, 9, 10 and 14 for the same swipe. Those reads came from the stale accessibility cache.
• This change: 15 of 15 passed on the old base and 10 of 10 on the rebased head. The anchor was always a fully visible row (10, 11 or 12) at its live position.

---

^{Lagoon commented through Stamp · server raft · head when posted 2aaccf0
View in Stamp →}