chore(toolpath-desktop): add synthetic preview benchmarks

[eliothedeman]

Summary

• Adds gen_synthetic_path binary in toolpath-cli that emits a synthetic Path JSON at configurable step counts (default --steps 1000), with a realistic ~70% text turns / ~20% Edit or Write / ~10% MultiEdit mix. Deterministic seed.
• Adds crates/toolpath-desktop/frontend/src/lib/__bench__/preview.bench.ts, runnable as bun run bench, covering the Preview's pure-TS hot paths: normalize, buildTree, flattenChatHead, classify, matchesFilter (keystroke simulation).
• Adds crates/toolpath-desktop/frontend/BENCHMARKS.md — describes how to generate fixtures, run the bench, and the manual Chrome DevTools procedure for render-time + memory (the parts a Tauri webview dominates and an agent can't script). Includes a 2026-04-22 baseline table with real pure-TS numbers and empty rows for the manual measurements.
• Bumps toolpath-cli 0.3.1 → 0.4.0 (additive public change: new binary). Workspace root Cargo.toml does not list toolpath-cli as a workspace dep, so only Cargo.toml + site/_data/crates.json + CHANGELOG.md needed updates per CLAUDE.md's checklist.
• .gitignores /bench/fixtures/ — fixtures are regenerated locally, not committed (~5 MB at 10k steps).

Addresses #41. Not Closes — the issue asks for a rerun after #38 and #39 land, so it stays open as a tracker.

Baseline numbers (Apple M4 Pro, bun 1.3.5)

Pure-TS, 10 iterations each:

Size	JSON.parse	normalize	buildTree (median)	buildTree (p95)	keystroke filter	flattenChatHead	classify × all
1k	1.16 ms	0.23 ms	3.98 ms	7.5 ms	0.08 ms	0.23 ms	0.14 ms
5k	3.17 ms	0.79 ms	82.2 ms	113 ms	0.43 ms	1.20 ms	0.47 ms
10k	6.32 ms	2.15 ms	579 ms	1830 ms	1.12 ms	5.34 ms	1.49 ms

buildTree at 10k is well above the 200 ms keystroke budget from the issue. Expected — this is the thing #39 is meant to address.

Tauri webview measurements (TFP, keystroke DOM-updated, memory delta) are left blank in BENCHMARKS.md with a manual procedure — those need a human at a running cargo tauri dev session with DevTools open.

Test plan

• cargo build -p toolpath-cli --bin gen_synthetic_path
• cargo run -p toolpath-cli --bin gen_synthetic_path -- --steps 1000 --out bench/fixtures/synthetic-1k.path.json (and 5k / 10k)
• cargo run -p toolpath-cli --bin path -- validate --input bench/fixtures/synthetic-1k.path.json (passes)
• cd crates/toolpath-desktop/frontend && bun install && bun run bench (produces the numbers above)
• bun run check (0 errors, 4 pre-existing warnings)
• bun run build (clean)
• cargo clippy -p toolpath-cli --bin gen_synthetic_path -- -D warnings (clean)
• cargo test -p toolpath-cli --bin path (152 pass, 0 fail)
• Maintainer runs the manual DevTools procedure and fills the Tauri rows in BENCHMARKS.md
• Rerun bun run bench after toolpath-desktop: memoize markdown rendering per chat turn #38 and toolpath-desktop: buildTree re-normalizes on any preview mutation #39 merge; append the comparison rows

[eliothedeman]

Summary

Solid execution of the benchmarking brief: a working fixture generator, a pure-TS bench that actually produces comparable numbers, and a BENCHMARKS.md that is honest about what was measured vs. left for manual. Scope stays clean — no perf fixes smuggled in, no svelte-virtual-list, no speculative sub-issues filed.

Notes

• crates/toolpath-cli/src/bin/gen_synthetic_path.rs:293 — real tool, deterministic seed, clap CLI, writes valid Path docs (PR confirms path validate passes). Mix (~70/20/10) matches the ask.
• .gitignore:5 — fixtures correctly gitignored under /bench/fixtures/, not checked in. Good.
• crates/toolpath-cli/Cargo.toml:51 — default-run = "path" preserves cargo run -p toolpath-cli behavior; second [[bin]] added cleanly.
• crates/toolpath-desktop/frontend/src/lib/__bench__/preview.bench.ts:574 — plain performance.now() + median/p95/max over 10 iterations. No Vitest dep added, which fits the minimal-surface-area vibe. KEYSTROKE_QUERIES cycles realistic prefixes.
• BENCHMARKS.md:489 — baseline table has real numbers for pure-TS ops; Tauri rows left as — with an explicit manual procedure. Matches the PR body's honesty claim.
• BENCHMARKS.md:509 — narrative correctly attributes the 10k buildTree regression (579 ms median, 1.8 s p95) to #39 territory and flags renderMarkdown as the likely keystroke culprit for #38, without doing the fix here.
• BENCHMARKS.md:551 — explicit "don't file sub-issues for known wins from #38/#39" — respects the scope-creep guardrail from the issue.
• Version bumps: toolpath-cli 0.3.1 → 0.4.0 in Cargo.toml, Cargo.lock, CHANGELOG.md:19, site/_data/crates.json:84. Workspace root [workspace.dependencies] does not list toolpath-cli (confirmed — it's a leaf binary crate), so that's correctly skipped. Checklist followed.
• tsconfig.json:765 excludes __bench__/** from svelte-check — sensible; the bench uses Node APIs.
• Minor: bench/fixtures/synthetic-10k.path.json TFP row in the manual procedure (step 3) candidly admits the paste-into-DevTools route is awkward and recommends using a real long session instead. Honest, slightly hand-wavy — fine as a starting procedure.
• No CI checks reported on the branch (gh pr checks 47 → "no checks reported"), but the PR body lists the local verification matrix.

Verdict

Approve. Delivers exactly the benchmarking scaffolding the issue asked for, with no scope creep and honest accounting of measured vs. pending numbers. Safe to merge; issue stays open as the rerun tracker post-#38/#39.

[github-actions]

🔍 Preview deployed: https://25cf3266.toolpath.pages.dev