fix(frontend): stop reduced-motion transcript guards from compounding / 修复关闭窗口动画后会话滚动大跳

[SivanCola]

Summary

User report: on Win10 LTSC 1809 (8th-gen i7, GTX 1060) with all Windows visual effects disabled, the transcript still jumps by thousands of pixels while scrolling and the question-jump slider is unusable. Enabling "Animate controls and elements inside windows" makes both symptoms disappear.

That setting is what WebView2 maps to prefers-reduced-motion: reduce (the attached diagnostics manifest has reducedMotion: true). Our global reduced-motion reset turned every transition into a 0.01ms transition: all. A near-zero transition still starts from the old value, so transform/padding writes are not visible to same-frame getBoundingClientRect / scrollHeight reads; they land one frame later. Verified in headless Chromium: normal mode reads the new value synchronously, the 0.01ms rule reads the old value until the next frame, transition: none reads the new value synchronously again.

The transcript scroll guards measure geometry right after their own writes, so:

• The reader-extent guard wrote its visual-guard transform, re-measured the anchor row, saw "no change", and treated its own unapplied transform as fresh displacement. The diagnostics show reverseDisplacement 7088.5 → 14177 → 21265.5 → 28354 → 35442.5 (exact multiples), followed by a correction to scrollTop 0. That is the reported jump.
• Virtuoso's padding writes also lagged a frame, so scrollHeight oscillated 9057 ↔ 16369 during the history prepend and the question-jump intent ended degraded. That is the "slider unusable" report.

Changes

• styles.css: the reduced-motion block uses transition: none !important instead of transition-duration: 0.01ms. Animations keep the 0.01ms squash so fill-mode: forwards end states survive.
• transcriptReaderExtentStability.ts: new transcriptAppliedVisualOffset() reads the item list's applied computed transform (matrix/matrix3d translateY) and falls back to the remembered offset only when nothing can be read.
• useTranscriptReaderExtentStability.ts (observe + tick correction) and transcriptAnchorCompensation.ts subtract the applied offset rather than the remembered one, so a lagging or externally cleared guard cannot compound.
• Race tests: two new scenarios (transform not yet applied; transform applied but remembered offset gone). Both fail against the previous hook (-2043px guard, 1154px correction) and pass now (-681px, 681px).
• Bundle budgets step one decimal each, measured on the rebased tree with the worktree's own install: gzip 462.827 KiB (budget 462.7 → 462.9), raw 2469.815 KiB (budget 2469.4 → 2469.9).
• Follow-up d8840832f: the two visual-guard race scenarios live in transcript-reader-visual-guard-race.test.tsx (the extent race file was back over the 800-line repolint ceiling), and check-motion-ci-contract.mjs now asserts the global reduced-motion reset uses transition: none !important and never shortens transitions, so the CSS half of the fix cannot regress silently.

Verification

• pnpm test (251 suites), typecheck, test:typecheck, lint:hooks, check:css, test:motion, pnpm build all pass.
• bench/transcript-scroll-stability.mjs with reducedMotion: "reduce" emulation: 94/94 pass, including the reduced-motion tail and return-to-bottom gates.
• bench/transcript-reader-transaction.mjs: Chromium 20/20 rounds pass; the WebKit leg did not run locally (WebKit not installed).
• transcript-reader-visual-guard-race: 9 pass on the head; 2 fail against the pre-fix hook. The motion contract check fails when the reset is reverted to transition-duration.
• go run ./tools/repolint clean.
• Not yet verified on the reporter's Windows machine; a canary test build from this head is with the tester.

Merge order

Merge this PR before #9754. Both touch useTranscriptReaderExtentStability.ts (adjacent hunks in observe()) and check-bundle-budget.mjs; a dry-run merge in either order conflicts on those two files, so #9754 should rebase onto main-v2 after this lands. The changes are semantically independent: this PR replaces the remembered visualOffset with the applied transform when deriving physicalAnchorDrift; #9754 widens the rejected predicate two lines below.

Documentation-impact: none - behavior change is a bug fix in the scroll guards and a reduced-motion CSS reset; no documented user-facing behavior changes.