fix(website): homepage interaction wedge refresh (B2 follow-up)

[blove]

Summary

Lands the PR #131 sort + filter comparator wedge on the homepage. Three editorial surfaces touched; no source/package changes.

No auto-merge. Three editorial surfaces means prose review — please read the prose drafts below and confirm before merging.

What changed

• `apps/website/app/components/ComparisonTable.tsx` — three new interaction rows between `scroll anchor shift (px)` and `headless engine + React surface`. Section subhead gets one new sentence. Header docblock cites the new milestone source.
• Trail-marker labels rewritten on three comparators (`pretable` unchanged).
• `apps/website/app/bench/page.tsx` — placeholder Interactions section replaced with a real section paralleling the H1 scroll layout (loader + verdict helper + table + two prose paragraphs).
• `scripts/extract-interaction-summary.mjs` + `status/milestones/2026-05-10-b2-sort-filter-summary.json` — one-shot aggregator + the milestone summary it produces from PR feat(bench): B2 follow-up #5b — sort + filter comparator wiring #131's per-run JSONs.
• `apps/website/tests/components/ComparisonTable.test.tsx` — regex regression-guards for the new trail-marker labels.
• `docs/research/repo-memory.md` — 2026-05-11 entry.

Comparison table — new interaction rows

metric	pretable	AG Grid	TanStack	MUI X	budget
sort latency p95 (ms) — interaction	16.5	58.3	34.4	35.0	≤ 16
filter-metadata latency p95 (ms)	16.0	49.9	15.7	33.4	≤ 16
filter-text latency p95 (ms)	17.7	50.0	40.2	33.3	≤ 16

Trail-marker label diff

Adapter	Before	After
pretable	Recommended path	unchanged
AG Grid	Slower scroll; row-height drift	1.7× slower scroll, 3× slower interaction; row-height drift
TanStack	Headless; you wire selection and nav	Headless; ~2× slower interaction (filter-metadata ties pretable)
MUI X	Parity at scroll p95; full-grid feature surface	Scroll-p95 parity; 2× slower interaction

The MUI X label is the most consequential change — the prior "parity at scroll p95" framing read positively about MUI; the new label preserves that half of the story while adding the interaction caveat.

ComparisonTable subhead — new closing sentence

Wrapped-text scroll at 3,000 rows on Chromium, against the three most-cited React grids. Pretable matches MUI at the front, ~1.7× ahead of AG Grid Community and TanStack on raw frame p95 — and is the only adapter here that clears every quality threshold (zero blank gaps, zero anchor shift, ≤1 px row-height drift) at full-grid feature weight. Interactive sort and filter run 2–3.5× faster than every measured comparator on the same dataset.

/bench page Interactions section — prose draft

Pretable sorts and filters 3,000 wrapped-text rows in 16–18 ms across all three scripts — clear of the single 60Hz frame budget on `filter-metadata` and `sort`, fractionally over on `filter-text`. AG Grid Community runs sort and filter 3–3.5× slower despite being a full feature-surface grid; MUI X DataGrid Community lands at roughly 2× across all three scripts. TanStack Table v8 + TanStack Virtual is the only comparator that ties pretable on a single metric — `filter-metadata` at 15.7 ms vs 16.0 ms, within run noise — but is 2.1× slower on sort and 2.3× slower on `filter-text`.

Like the scroll story, the H6/H7/H8 evaluators check pretable's absolute thresholds (`≤ 32 ms` interaction p95) rather than gating on comparator parity. All three hypotheses stay satisfied at n=3.

What's NOT in this PR

• `ReceiptsBand.tsx` — PR fix(website): reframe streaming claims as capability-anchored after S5 cross-validation #129 (streaming reframe) is still open and modifies this file. Left untouched here to dodge a merge conflict over an unresolved editorial decision.
• `FeatureGrid.tsx` — already capability-anchored after PR fix(website): retire stub-era "4× faster" claims; reframe homepage around B2 evidence #126; doesn't need a touch.
• High-repeat (n=20) follow-up for pretable's `filter-text` at 17.7 ms (1.7 ms over budget) and TanStack's `filter-metadata` at 15.7 ms (the tie). Both within ±2 ms of the frame budget; a follow-up could clarify but isn't a v1 blocker.
• Comparator-aware H6/H7/H8 evaluators — the page reads from the aggregated summary file directly; evaluator-array extension is future work.

Test plan

• `pnpm -w typecheck` passes
• `pnpm -w test` passes (190 website tests; trail-marker regex regression-guards updated)
• `pnpm -w lint` 0 errors
• `pnpm format` clean

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
pretable	Ready	Preview, Comment	May 11, 2026 4:31pm

[github-actions]

Vercel preview ready

Preview: https://pretable-cdqzdb5nx-cacheplane.vercel.app
Commit: 25ecc88b6168aa44c8726e24dad06081078508fe

_{Updated automatically by the deploy-preview job.}