fix(ui): returning to a live conversation intermittently re-streams accumulated output — regression from #2975, flaking CI #3044

Symptom

streaming-remount.spec.ts:65 ("returning to a live conversation settles output accumulated while away") fails intermittently on main: after switching back to a backgrounded live conversation, the output that accumulated while away is replayed progressively (partial-text paints) instead of appearing settled. This is a user-visible product regression, not only a flaky assertion: the transcript visibly re-streams content the model already produced.

Observed failures on main CI: run 31791670100 (2026-08-14). It has also hit unrelated PRs twice (#3032).

Bisect

Locally reproducible at high rate on macOS (Darwin 25.6.0, Node 24.18):

Evidence

Instrumenting the spec's MutationObserver shows the streaming bubble's text after switching back growing in small per-paint steps until it reaches the full accumulated text (fail run):

"…stop the Turn."
"…stop the Turn.Acknowle"
"…stop the Turn.Acknowledged steer"
"…stop the Turn.Acknowledged steering: backg"
"…stop the Turn.Acknowledged steering: background outp"
… (ends at the full string)

The fake backend emits the steering echo as a single text_delta, and the projector's seedActive emits accumulated text as one cumulative delta — so these partials are not source chunks; the catch-up is being applied/painted incrementally after the remount.

Two mechanisms interact, both new or newly-exposed in #2975:

1. app-shell-session-events.ts now defers active-session text_delta/thinking_delta application to a per-frame flush (scheduleDisplayEvent, rAF + 100ms fallback). A catch-up burst that spans multiple tasks/frames is committed in several partial paints.
2. The virtualizer mounts the streaming turn (and paints) earlier on session switch than the previous progressive-mount path did, so the first paint now lands before the catch-up burst has finished arriving — pre-perf(ui): virtualize long chat transcripts #2975 the burst had effectively completed before the streaming bubble mounted, which is why the same spec passed.

Additional diagnostic: holding a second persistent window.maka.sessions.subscribeEvents(sessionId, …) subscription open across the switch makes the failure deterministic and total — the returning view then never receives the accumulated echo at all (bubble stuck at the pre-steer text past a 10s timeout). That experiment perturbs observer registration, but it points at the catch-up/seed path racing observer lifecycle rather than at paint timing alone.

Expected

Switching back to a live conversation renders everything that accumulated while away settled in the first paint (the spec's contract); only genuinely new output streams.

Notes

• Reproduce: cd apps/desktop && npm run build:with-deps && npm run build:test && npx playwright test e2e/streaming-remount.spec.ts:65 --config e2e/playwright.config.ts --repeat-each 4.
• A fix probably wants the session-activation path to drain/apply the pending catch-up for the newly active session atomically before the streaming bubble's first paint, rather than letting it trickle through the per-frame display batch — but that touches perf(ui): virtualize long chat transcripts #2975's frame-batching design, so deferring the approach to its author.

Bisect and analysis by Claude Code.