feat(bench): opt-in CDP tracing in bench.spec.ts (unblocks flame-graph perf diagnostics)

[blove]

Summary

• Adds an opt-in Chrome DevTools-format trace capture path to apps/bench/tests/bench.spec.ts, gated by PLAYWRIGHT_PERF_TRACE=1. Output: status/traces/<stem>.cdp.json (sibling to the Playwright .trace.zip, gitignored).
• Closes the tooling gap surfaced by PR docs(research): pretable wrapped-text filter perf diagnostic #142's wrapped-text filter perf-diag: Playwright's default tracing.start({ screenshots, snapshots }) produces an SDK action trace, not a JS flame graph. CDP Tracing.start over disabled-by-default-devtools.timeline + v8 + disabled-by-default-v8.cpu_profiler produces DevTools-loadable JSON.
• Best-effort additive: CDP attach failures, end timeouts, JSON write failures all log + continue. Env-unset path is byte-identical to current behavior — the existing spec test still passes unchanged.

Manual verification

• Captured one CDP trace locally for pretable / S2 / hypothesis / filter-text. File: status/traces/chromium-pretable-default-s2-hypothesis-filter-text-2026-05-13t17-55-56-984z.cdp.json (30 KB, 145 events).
• jq confirms well-formed shape: every event has ph, ts, cat, name. Categories present: __metadata, disabled-by-default-devtools.timeline, disabled-by-default-v8.cpu_profiler, v8 — the standard DevTools profiling set, as expected.
• DevTools-load verification deferred to the reviewer. The captured JSON is in DevTools trace-event format ({ traceEvents: [...] }) and should load in Chrome → DevTools → Performance → "Load profile…".

Known limitation (consumer side, not the harness)

The bench app starts running its interaction script on page-load. CDP attaches after the adapter-label visibility check, so the very first moments of the interaction may run before tracing starts. For hypothesis-scale runs the captured window is sparse (~145 events). The wiring + format are correct; the next perf-diag consumer (the deferred wrapped-text filter investigation) will likely need either a wider scale or a triggered-start variant in the bench app. Documented in the spec, docs/research/repo-memory.md, and called out as out-of-scope here.

What's NOT in this PR

• Matrix-runner integration (would multiply matrix wall-clock 5–10×).
• Speedscope export.
• Automated test for the CDP path (requires a browser-in-CI setup beyond what we have; manual verification is the bar).
• Wrapped-text filter perf-fix v2 (separate follow-up; this PR unblocks the tooling).

Files touched

• apps/bench/tests/bench.spec.ts (+62)
• docs/research/repo-memory.md (+28 / -1)
• docs/superpowers/specs/2026-05-13-bench-harness-cdp-tracing-design.md (new)
• docs/superpowers/plans/2026-05-13-bench-harness-cdp-tracing.md (new)

No `packages/` changes. No public-API surface.

Gates

• `pnpm -w typecheck` — passes
• `pnpm -w test` — 191 tests pass
• `pnpm -w lint` — passes
• `pnpm format` — clean
• Manual verification: CDP JSON well-formed (`jq` shape check + category enumeration)

🤖 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 13, 2026 6:14pm

[github-actions]

Vercel preview ready

Preview: https://pretable-p57kyj3hj-cacheplane.vercel.app
Commit: 36cc0b21a328b7ac643422e0006b58785db85c1d

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