fix(desktop): stop at-rest scroll jump-up during code-block highlight

[teknium1]

Summary

Desktop chat no longer jumps up and snaps back while reading at the bottom — the symptom that survived #38221. Follow-up fixing a third, independent scroll-anchor bug on the armed-at-bottom path that #38221's two disarm fixes didn't cover.

Root cause (verified live in headless Chromium over CDP, not just modeled): while parked at the bottom, Streamdown/Shiki re-tokenizing a code/patch block briefly replaces laid-out DOM, so for one frame the content is shorter. The browser clamps scrollTop upward the instant content shrinks below the scroll position (1400 → 1100 with no pin involved); the next frame content regrows and the rAF pin snaps it back down. That's the "jumps quickly upward then returns" bounce, reported on Win11/macOS after updating, and one reporter pinned it to code blocks.

A pin cannot prevent it — the clamp happens at layout, before any pin runs (confirmed: both deferred and synchronous re-pins still show 1400 → 1100 → 1480).

Fix

Keep the content's height monotonic within a turn. The ResizeObserver raises a high-water-mark of the content height and reserves it as min-height on the content before pinning, so a transient shrink never shrinks the scroller and the browser never clamps. scrollHeight still grows under the viewport, so streaming tokens follow normally. The high-water-mark is reset in jumpToBottom (new turn / session / first content) and on user disarm, so an old tall thread can't pad a new short one and a finished turn can't leave a dead gap.

Changes

• thread-virtualizer.tsx: contentHwmRef + min-height reservation in the RO callback; reset in jumpToBottom and disarm.
• streaming.test.tsx: regression that models the browser clamp-on-shrink (scrollHeight = max(measured, reserved min-height), scrollTop clamps on shrink). Armed at bottom, a shrink RO frame must keep scrollTop at 1400 and reserve min-height: 2000px.

Why #38221's tests missed it

Both #38221 regression tests exercise the disarmed / scrolled-up path. This bug is on the armed / parked-at-bottom path — a disjoint code path — and the root cause is the browser's own clamp-on-shrink, which jsdom doesn't model.

Validation

	Before (main)	After
live CDP, shrink frame while armed	scrollTop 1400 → 1100 (up-jump) → 1480 (snap back)	1400 → 1400 → 1400 (no bounce)
streaming follows growth	yes	yes (unchanged)
streaming.test.tsx	new test RED (min-height stays empty)	11/11 pass
tsc -b	—	0 errors
eslint (changed files)	—	0 errors (only pre-existing padding-line / react-compiler warnings)

Live proof was a headless google-chrome + CDP harness driving a real scroller through the shrink→grow frames with native scrollTop clamping — the honest "real browser" bar for a renderer scroll bug, since jsdom doesn't clamp.

Known tradeoff

A legitimate mid-turn shrink (a tool-call panel collapsing, the loading indicator removed at turn end) is held at the high-water-mark until the next jumpToBottom/disarm, which can briefly reserve a few px of empty space at the very bottom. This resets on the next user message / thread.runStart / session change. Far less disruptive than the bounce, and only while armed at bottom.

Infographic

[github-actions]

🔎 Lint report: hermes/hermes-aba10f53 vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 9736 on HEAD, 9736 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 5045 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.

[tonydwb]

Code Review Summary

Verdict: Approved

The desktop scroll-anchor logic now defends against the Shiki/Streamdown transient-shrink bug (the code-block highlight phase briefly replaces laid-out DOM with shorter content, causing the browser to clamp scrollTop upward). The fix:

1. Tracks a contentHwmRef (high-water mark) reset per turn
2. Reserves min-height = high-water-mark during the armed state via ResizeObserver
3. Clears min-height on disarm and jumpToBottom so prior thread content doesn't pad new sessions
4. Guards the virtualizer's scrollToFn against fighting the pinToBottom loop (sticky-bottom ref)
5. Key fix for IME: event.nativeEvent.isComposing guard on Enter

✅ Looks Good

• Fixes the root cause (monotonic height) rather than masking symptoms
• Tests are thorough: transient-shrink, idle layout shift, viewport resize during streaming, IME guard
• IME composition guard is essential — pressing Enter to confirm a composed word must not submit

💡 Suggestion

• stickyBottomRef is a shared mutable ref between useThreadScrollAnchor and the VirtualizedThread virtualizer config. This coupling is correct but worth documenting in a comment near the virtualizer's scrollToFn since it crosses two abstraction boundaries.

---

Reviewed by Hermes Agent