feat(bench): interaction-window slicing for CDP traces + correction to PR #145

[blove]

Summary

• Adds three `performance.mark` calls to bench-runtime (`pretable.interaction.start`, `.firstFrame`, `.settled`) and `blink.user_timing` to the CDP categories list, so captured traces include the interaction window bounds.
• Extends `scripts/analyze-cdp.mjs` with `--window=interaction|settle|full`. Default remains `full` for backwards compat.
• Corrects PR docs(research): wrapped-text filter perf-fix investigation — hotspot identified, fix queued #145's misattribution. With the new slicing, `getEstimatedRowHeightSignature` (F8) does not appear in the interaction-window top-26 hotspots — its full-trace cost was during initial mount (3000-row cache fill), not the filter trigger.

What the sliced trace shows (n=1)

Window: interaction (6.97 ms)
Total sample time in window: 5.26 ms across 26 unique nodes

Top 5 by SELF time (sourcemap-resolved):
  1127μs  (21.4%)  (program)                                  native overhead
   802μs  (15.3%)  matchesFilters  packages/core/dist/index.mjs:69
   409μs  ( 7.8%)  (anon) packages/react/dist/index.mjs:29
   275μs  ( 5.2%)  u8 packages/react/dist/index.mjs:17
   272μs  ( 5.2%)  Hd react-dom-client.production.js:13346

The real interaction-window hotspot is `matchesFilters` (`packages/grid-core/src/derived-rows.ts:73`) — lowercasing every cell value on every filter call across all 3000 source rows. ~800μs of the 6.97 ms window.

What's NOT in this PR

• The actual `matchesFilters` perf fix (lowercase cache). Queued as a clean follow-up; should move filter-text from 16.79 ms p95 (PR docs(research): interaction borderline perf diagnostic (real over-budget on pretable filter-text) #134's n=20) to ~16.0-16.2 ms — borderline under budget.
• Sort + filter-metadata slice (separate trace captures).
• Matrix-runner CDP integration, Speedscope export — same status as PR feat(bench): opt-in CDP tracing in bench.spec.ts (unblocks flame-graph perf diagnostics) #143.

Lesson saved to memory

"Always slice CDP traces to the interaction window before naming a hotspot — full-trace view is dominated by mount-time work that doesn't count against `interaction_latency_ms`."

Gates

• `pnpm -w typecheck` — passes
• `pnpm -w test` — 191 / 191 pass
• `pnpm -w lint` — passes
• `pnpm format` — clean
• No `packages/` source change. Quality wedge untouched.

🤖 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 18, 2026 7:33pm

[github-actions]

Vercel preview ready

Preview: https://pretable-d7v8igkli-cacheplane.vercel.app
Commit: bbdf576641f9eb2ad9a7aced0f6a1a9f7dfd69c9

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