perf(cdp): add benchmark harness report

[aram-devdocs]

Summary

• Add Criterion benchmark suite (crates/plumb-cdp/benches/cdp_benchmarks.rs) with three groups: per_rule_dom (100/1k/10k node synthetic DOMs), cold_start, and warm_run (CDP benchmarks behind e2e-chromium feature gate)
• Wire just bench and just bench-full targets in the Justfile
• Add docs/src/performance.md with benchmark usage, fixture descriptions, and result interpretation guidance

Validation

• cargo fmt --all -- --check — clean
• cargo clippy -p plumb-cdp --benches -- -D warnings — clean
• cargo bench -p plumb-cdp --no-run — compiles successfully
• git diff --check — no whitespace issues
• CDP benchmarks (cold_start, warm_run) require a real Chromium binary and are gated behind the e2e-chromium feature; they will not run in CI without Chromium installed

Blockers

• No automated CI threshold gate (out of scope per bench: cold-start, warm-run, per-rule cost across DOM sizes #61)
• cold_start and warm_run benchmarks cannot execute without Chromium; the per_rule_dom group runs without it

Refs #61

🤖 Generated with Claude Code

[github-actions]

I have all the context I need to render the review.

---

PR #180 — perf(cdp): add benchmark harness report

Scope: crates/plumb-cdp/benches/cdp_benchmarks.rs (new), Cargo.lock, crates/plumb-cdp/Cargo.toml, docs/src/performance.md (new), docs/src/SUMMARY.md, justfile.

---

1. Determinism invariants — CLEAN

No library code changed. The bench binary calls plumb_core::run, which is the deterministic core; the harness itself is excluded from the determinism CI job. criterion pulls in oorandom (an RNG) but only inside the harness binary, not in any observable output path. No SystemTime::now or HashMap leak into library code.

2. Workspace layering — CLEAN

criterion is correctly placed under [dev-dependencies] in crates/plumb-cdp/Cargo.toml, referenced via { workspace = true } where the workspace root declares criterion = { version = "0.5", features = ["html_reports"] } at line 101. The bench binary sits in plumb-cdp (the one crate allowed to hold unsafe); the layering rule is unaffected.

All four public types used in the e2e-chromium-gated mod cdp block (BrowserDriver, ChromiumDriver, ChromiumOptions, PersistentBrowser, Target) are confirmed present in crates/plumb-cdp/src/lib.rs.

3. Error handling — CLEAN

unwrap / expect appear only inside the bench binary with a file-level #![allow(clippy::unwrap_used, clippy::expect_used)]. That allow is scoped to a standalone binary, not a library crate — no rule violation.

4. Dependency hygiene — CLEAN

New transitive deps (anes, cast, ciborium*, crunchy, half, hermit-abi, is-terminal, itertools 0.10.5, oorandom, plotters*, same-file, tinytemplate, walkdir, winapi-util) are all MIT / Apache-2.0 / BSD-2 or dual-licensed combinations, all within the deny.toml allow-list. itertools 0.10.5 has no prior version in the lockfile — no multiple-version warning.

5. Documentation — CLEAN

docs/src/performance.md is concise and factual. No AI-writing tells from the documented list ("dive in", "seamless", "leverage", "comprehensive", etc.) were found.

---

Punch list

• crates/plumb-cdp/benches/cdp_benchmarks.rs:106–113 — Three inline #[allow(clippy::cast_possible_truncation)] / #[allow(clippy::cast_possible_wrap)] are dead noise; the file-level #![allow(..., clippy::cast_possible_truncation, clippy::cast_possible_wrap, ...)] on line 23 already suppresses them. Not a blocker, but tidy-up is cheap.

---

Verdict: APPROVE