[yaacovcorcos]

Platforms: measured and tested in headless Chromium on macOS (Apple Silicon). Electron 44 uses the same engine on Windows and Linux, but neither was run.

What Changed

Follow-up to #801 and #806. Scrolling a Visual LaTeX document no longer re-lays out the paged document, and showing the LaTeX tab costs about half as much. All changes are in Scient files; no T3 file changes.

Scrolling (latexPagedPresentation.ts):

• Reveals restarted layout passes that changed nothing. A content-visibility reveal now schedules a pass only when the block's size actually changed (reported by the shared observeResize). Unknown blocks keep the old path.
• Passes never settled. Spacers recomputed from getBoundingClientRect drifted by up to 0.55 px from the published values, so about 400 of 768 decorations were replaced on every pass. A gap or offset within 1 px of the published one now keeps it.
• Every pass rewrote every block's intrinsic size, because a string comparison could never match. Sizes are now compared as numbers.
• Bug fix, also on main: a pass read block positions across frames, so a reveal between slices moved the blocks read after it. At 4× pages of text jumped by about 380 px. Positions are now read in one layout.

Showing the tab (new latexBlockNeighbours.ts):

• Five sibling :has(+ …) selectors made Chromium restyle every top-level block on each block the loader inserted.
• A ProseMirror plugin now marks a paragraph before a display (data-latex-before-display, "short" when the paged layout set the display short) and a run-in heading before its paragraph (data-latex-run-in). The CSS uses those attributes.
• After an edit, it rebuilds only the blocks around the change: 0.2–0.3 ms per key on a 783-block document.
• Guard: universalSiblingSelectors.test.ts now also rejects sibling :has() under src/scient.

One visible difference: a paragraph holding a footnote, followed by a short display, is now raised above the display like every other such line. Before, the note's band between them in the DOM hid the match, and the formula took clicks on that line's lower half. Computed styles of every paragraph, heading and display were otherwise identical across 11 editor states.

Why

61-page Visual LaTeX document, production build:

	main (with #806)	this PR
Scrolling at 4× CPU throttle, long frames per ~15 s	12–13 s	about 1 s
Typing while scrolling, unthrottled, p50/p90 per key	64/80 ms	16/16 ms
Showing the LaTeX tab at 4×, long frames (median)	6.1 s	2.9 s
:has() invalidation records per tab show	44,229	1,735

PDF scrolling needed nothing more: #806 already removed its cost.

What remains of tab show is mostly the real style and layout of the content. Keeping the editor mounted would remove it, but it needs T3 panel changes and 0.3–0.4 GB per thesis-sized document, so it is not done here.

Verification

• vp run check:changed origin/main passes: format, lint, knip, typecheck, 7 affected test files and the React Compiler check.
• All LaTeX browser suites pass: 53 files, 560 tests (run before the last review fix; the affected suites ran again after it).
• New latexBlockNeighbours.test.ts checks the plugin against a fresh one after an explicit sequence and 400 seeded random edits, undo/redo and layout publications. Deliberate breaks to the plugin make it fail, including the attribute-step case below.
• New latexPagedPresentation.browser.test.tsx cases cover the reveal gate and the single-layout read.
• Codex review (gpt-6.1-sol, high): approve with fixes. Fixed in the last commit:
  • a heading level change made through setNodeAttribute maps no positions, so the plugin missed it;
  • the guard missed a sibling selector later in a :has() list.
• Not run locally: the full test suites, the full build and the desktop smoke test. CI covers the suites and build.

Documentation Impact

Documentation impact: Updated — docs/internals/scient-latex-visual-performance.md (scroll causes and the remaining tab-show cost).

Checklist

• This PR is small and focused
• I explained what changed and why
• Visible change described above (footnote line above a short display)

Implementation: Claude Code (Claude Opus 5.5). Review: Codex (gpt-6.1-sol).

🤖 Generated with Claude Code

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ The exact PR base did not have a successful artifact. Baseline uses the latest successful main measurement shown below.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	5.5 KiB	5.5 KiB	0 B (0.0%)	6.8 KiB	✅
Codex	Thread snapshot wire	4.2 KiB	4.2 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.3 KiB	1.3 KiB	0 B (0.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	22.3 KiB	22.3 KiB	0 B (0.0%)	24.4 KiB	✅
Codex	Live turn messages	1	1	0 (0.0%)	8	✅
Claude	Total thread wire	5.5 KiB	5.5 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	4.2 KiB	4.2 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.3 KiB	1.3 KiB	0 B (0.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	22.7 KiB	22.7 KiB	0 B (0.0%)	24.4 KiB	✅
Claude	Live turn messages	1	1	0 (0.0%)	8	✅

Baseline: bfa08d9 · PR result: 829ad36 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.
Bounded compact HTTP snapshot with live cursor.

• Codex decoded thread snapshot: 88.2 KiB
• Claude decoded thread snapshot: 88.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.