flaky(e2e): streaming-remount:142 bubble mounts as historical without maka-bubble-streaming while the Turn is live

[me2seeks]

Symptom

e2e/streaming-remount.spec.ts:142 › keeps a completed reply after an interrupted turn and conversation remount intermittently times out (20s) waiting for:

await expect(page.locator('.maka-bubble-streaming')).toContainText('Fake backend waiting', ...)

element(s) not found — but the Playwright error-context snapshot shows the text is rendered (log > paragraph: "Fake backend waiting for the test to stop the Turn."). The answer bubble mounts with phase: 'historical' instead of the live streaming phase, so it never carries the maka-bubble-streaming class even though the sidebar shows the turn as 正在响应 (running).

So the race is in live-turn arming/projection: the new turn's first content lands through the historical transcript path while the Turn is actually live.

Reproduction (local, xvfb-run -a, --workers=1, --repeat-each=6)

Tree	Result
main @ 91624f1	1 fail / 6 runs
feat/4364-runtime-form-interactions @ b65d7ff	3 fail / 6 runs
feat/4364-desktop-form-interactions @ e8b22e8	1 fail / 3 runs

The flake exists on plain main; the #4364 stack may raise the rate but does not introduce it.

CI sightings

• PR feat(desktop): answer structured form interactions #4384 run 33746990744 job 100621633447 — spec:142 failure
• Same run, rerun job 100651199089 — different spec failed (skill-draft-lifecycle.spec.ts:110), also draft/remount-timing territory
• PR fix(desktop): publish MCP tools through native capabilities #4490 run 33739469134 — transcript-scroll.spec.ts:694 distance-58 flake (separate, reran green)

Related history

• flaky(e2e): unfenced Enter races a 30s send admission against a 10s expect (streaming-remount:161, transcript-scroll:316) #4573 (closed): unfenced Enter races send admission (streaming-remount:161)
• test(desktop): streaming-remount E2E intermittently misses the active Turn #3177 (closed): intermittently misses the active Turn
• fix(ui): returning to a live conversation intermittently re-streams accumulated output — regression from #2975, flaking CI #3044 / fix(desktop): do not replay streamed text after returning to a conversation #2895 (closed): re-streaming on conversation return

The 142 case looks like a cousin of #3177: the bubble loses the live classification on a freshly started Turn in a fresh conversation (the preceding step parks a draft in another session and clicks 新任务). Hypothesis: the first content event for the new Turn arrives before the renderer arms its live-turn projection for that session, so it is absorbed into the historical transcript and never re-classified.