Progressive transcript hydration exposes its moving watermark to unrelated live renders (vertical motion on resume)

[code-yeongyu]

Context

Found during the 2026-08-22 TUI vertical-jitter forensics (follow-up to #1064; the primary status-dock fix ships separately). This is an independent vertical-motion path for resumed / post-compaction transcripts, NOT the cause of the captured turn-boundary jitter.

Mechanism (HEAD 98234d1)

12d771dbc (perf(coding-agent): render resumed transcripts progressively) introduced progressive transcript hydration:

• packages/coding-agent/src/modes/interactive/components/progressive-transcript-container.ts:76-96 renders from hydratedFrom to the tail.
• :142-176 warms a head chunk per setImmediate and moves hydratedFrom BACKWARD after every chunk.
• Any unrelated assistant/tool render between chunks calls render() and exposes the newly lowered watermark, inserting a warmed batch ABOVE already-painted tail content -> painted content moves vertically in multiple geometry-changing batches while the user watches.

Affects resumed sessions and compaction rebuilds whose history exceeds the tail budget.

Proposed fix (preserves the perf intent)

Split warming state from visible state:

1. Keep warmedFrom for background cache population (Markdown work stays spread across bounded macrotasks).
2. Keep visibleFrom fixed at the initial tail boundary while warming is incomplete; unrelated live renders keep painting visibleFrom..total including appended live children.
3. When warmedFrom reaches zero, atomically set visibleFrom = 0 and issue the existing single completion render (one atomic height increase).

Failing-first proof

Deterministic case for packages/coding-agent/test/suite/progressive-transcript-container.test.ts:

1. Populate more children than tailBudget.
2. Render once; record the first visible component index.
3. Await exactly ONE setImmediate (one warm chunk completes, hydration unfinished).
4. Append one live component; render again.
5. Assert the first visible index is unchanged and the live component is at the tail.

Current HEAD fails: warmNextChunk() has already lowered hydratedFrom, so the second render starts one chunk earlier. No sleeps needed.

Risk notes

Medium: the final full-history reveal remains one atomic height increase and may still trigger one main-screen scrollback replay; verify resume, compaction rebuild, append-during-hydration, clear/dispose cancellation, and full transcript output.