Chat webview: one sweep per frame instead of one sweep per streamed chunk

While a task runs, the chat webview (src/kiss/agents/vscode/media/main.js, used identically by the VS Code extension and the remote web app) receives a stream of events. The overwhelming majority are tiny text chunks: thinking_delta, text_delta, and system_output (live bash output). A long build or test run delivers dozens of these per second.

The problem

The chunk's text was already batched: it goes into a buffer and is written to the DOM once per animation frame (a pending requestAnimationFrame "flush" that also auto-scrolls the panels it grows). But processOutputEvent() still ran a synchronous "tail" after every single chunk, and every part of that tail walks the whole transcript or forces a layout pass:

Per-chunk tail work (before)	Cost per chunk
`collapseOlderPanels()` via `streamEnd()` — collapse every top-level panel but the last (ran for every bash output chunk)	O(panels) scan + class churn
`autoScrollLatestEventPanel()` — `querySelectorAll` over the latest panel, then `scrollHeight`/`clientHeight`/`scrollTop` reads and writes	forced reflow
`applyChevronState()` — `O.querySelectorAll('.collapsible')` plus a `closest('.adjacent-task')` per panel	O(panels) DOM scan
`updateVisibleTask()` (and `showLiveMetrics()`) — full-tree selector scans	O(transcript)

All of this ran against layout that had not changed yet — the text was still sitting in the buffer waiting for the flush frame. With N panels in the transcript and a chunk stream, the tail costs O(N) per chunk and O(N²) over a long task: exactly when a task gets long and busy, every additional chunk gets more expensive.

The fix: defer the tail into one per-frame sweep

Per-chunk O(N) tail work (red) is replaced by a single per-frame sweep (yellow) that runs after the existing text flush (green) and before the browser paints — so nothing the user can see changes, only when the bookkeeping runs.

A chunk takes the fast path only when it is provably invisible until the next frame: its DOM write is buffered behind a pending flush (thinkRaf/txtRaf/bashRaf) and it did not open or adopt a thoughts panel (streamBegin counted no step). Everything else — panel-opening chunks, plain system_output without a bash panel, tool_call/tool_result/result/usage_info, background-tab events, and transcript replays — keeps the exact synchronous path it had before. Test environments that stub requestAnimationFrame to run synchronously are automatically unaffected: their flush handle is 0, so every chunk stays on the old path.

Independent review (gpt-5.6-sol, read-only)

A second model reviewed the change against the repo with instructions not to invent problems ($15.14 of a $30 cap, under the 50%-of-budget limit). It confirmed the skip conditions and found 5 real defects; all are fixed and each fix is pinned by a new test:

#	Severity	Finding	Fix
1	High	A sweep still pending when the task ends ran with `isRunning=false` and hid the finished task's panels (`chv-hidden`), dropping collapse debt.	`flushStreamTailSweep()` inside `setRunningState(false)`, before the flip.
2	Medium	A stale user scroll lock was no longer cleared: the collapse now happened after the flush with no autoscroll pass behind it.	The sweep runs `autoScrollLatestEventPanel(O.lastElementChild)` after collapsing.
3	Low	Sibling subpanels of the latest panel (e.g. a finished `.think`) were no longer pinned to their end on deferred flushes.	Same fix as #2 (that call scrolls every subpanel).
4	Medium	The new coverage gate copied a pre-existing off-by-one from the sibling gates (`n < b - 1`) that silently skips each region's last line.	Fixed to `n < b` in the new gate; the sibling gates' copies are noted as pre-existing and left untouched.
5	Low	A sweep retargeted across a tab switch kept its old rAF queue slot and could fire before the new tab's flush.	Retargeting cancels and re-arms the rAF behind the new tab's flush.

Verification

New suite test/streamTailCoalesce.test.js — 10 end-to-end jsdom scenarios (extension and remote webview): chunk deferral and per-frame replay, panel-opening chunks staying synchronous, thinking/text/bash parity of text and scroll state, synchronous-rAF fallback, tab-switch debt drop, retarget-on-first-chunk-after-switch, task-end flush (no chv-hidden ever), sibling-subpanel scroll parity, replay cancellation. Two defensive branches that cannot occur without mocks are documented in the test header instead of being faked.
New coverage gate test/streamTailCoalesce.coverage.js — requires and gets 100% line coverage (116/116) of every streamtail-coverage region in main.js.
Full regression: all 233 webview test files pass (run as 30 parallel splits), including the pre-existing 100% gates for auto-scroll, visible-task, cross-tab leak, chevron and timestamp behavior.
npm run lint:ts, lint:css, typecheck, and the repo-wide uv run check --full all pass.

Files changed (staged with `git add`, not committed)

src/kiss/agents/vscode/media/main.js — +133/−1 lines: sweep machinery (scheduleStreamTailSweep, runStreamTailSweep, flushStreamTailSweep, cancelStreamTailSweep), the deferTail fast path, and the three ordering hooks.
src/kiss/agents/vscode/test/streamTailCoalesce.test.js — new e2e suite.
src/kiss/agents/vscode/test/streamTailCoalesce.coverage.js — new coverage gate.