Fix long thread timeline scroll stalls

[SawyerHood]

Summary

• replace per-frame linear timeline geometry scans with debounced logarithmic lookups using fresh row geometry
• bound scroll-anchor capture work to top-level timeline rows while preserving prepend and streaming anchoring
• avoid rendering the closed TOC panel and stabilize outline-derived inputs during streaming updates
• preserve keyboard, pointer, focus, and overflow affordance behavior for the thread table of contents

Profiling

Before, the fully loaded long thread produced 33 long tasks, 4.21s aggregate blocking, a 190ms worst task, and 32 frames over 50ms during a bottom → top → bottom sweep. After removing the forced-layout hot paths, the same thread produced 0 long tasks and no frames over 50ms with flat heap. A synthetic 1,100-row sweep remained frame-smooth, so virtualization was not warranted after the measured low-hanging work was removed.

Validation

• pnpm exec turbo run typecheck --filter=@bb/app
• full @bb/app suite: 295 files / 2,169 tests
• post-rebase focused suite: 32 tests
• live browser QA for bottom anchoring, keyboard/pointer access, focus preservation, overflow fades, and geometry-read count
• independent GPT-5.6-Sol review: APPROVE