[tracking] filter-work-conversations Storybook smoke failed on three unrelated PRs within one night

[Astro-Han]

What this tracks

On 2026-09-18/19 UTC,unrelated PRs each failed a Storybook smoke step on a component they did not touch. Three of the five failures below belong to the same filter-work-conversations story family, across three unrelated PRs. Each single failure looks like flake; three hits on the same family in one night suggests the story itself, or the render path it depends on, may be timing-sensitive. No common cause is claimed here — this issue only collects the data so the pattern is not lost.

Instances (all verified against exact head SHAs)

#	PR	head	run / job	failing step / assertion
1	#5489	d8a00a0f	run 35355625325 / job 105634319820	product-workhub--filter-work-conversations-narrow
2	#5481	3c8c6490	run 35361965779 / job 105655159902	filter-work-conversations-narrow ([data-search-highlight] null)
3	#5499	934eeb9a	run 35415116847 / job 105822123031	filter-work-conversations (wide, expected length 4 got 1)
4	#5429	0799593b	run 35359768471 / job 105647974519	appshell--dock-affordance-returns-to-tail (expected false to be true)
5	#5489	57df2e4e	run 35412341506 / job 105814277531	WorkHub native-click smoke ('Test' !== 'Clicked')

All five: attempt=1, conclusion=failure. Job links: https://github.com/apache/maka/actions/runs/<run>/job/<job>.

Evidence in both directions (both must stay in this issue)

• For environment/timing: on the fix(ui): mount transcript rows far enough ahead for the tallest Turn #5499 head, the same WorkHub story passed 10/10 consecutive local reruns in the same worktree, and the base/main run was fully green. On that head it is not reproducible.
• Against pure flake: the same story family was hit three times (Add Rive workflow Maka tool #1, Harden runtime, storage, gateway, credentials, and IPC inputs #2, Extract runtime kernel boundaries and durable AgentRun recovery #3) on unrelated code. Rerun-passing on one head does not explain why the family keeps getting picked.

Known evidence gap

Instance #2 can never be re-judged on its original SHA: its attempt 2 was cancelled by a newer push and 3c8c6490 has since been rebased away. Do not assume it can be rerun.

Suggested next step (for CI owners)

Investigate whether the filter-work-conversations story (narrow and wide) or its shared render path is timing-sensitive under parallel Storybook smoke runs, rather than attributing each hit to random flake. If the story needs a retry/quarantine, that decision should be explicit, not silent.

---

Automated review notice: This issue was filed by an automated review agent operated by Astro-Han. It is not an independent human review and does not replace one.