path query: parallel per-file execution (path-cli 0.17.0)

[benbaarber]

Stacked on #136. Part of the query-acceleration work.

What

path query now evaluates cache documents on a thread pool:

• PerFileStream / Decompose plans treat every document independently, so the whole per-file pipeline — read, parse, wrap, filter, render/pack — runs on rayon workers (query/mod.rs::for_each_file, chunked). Only ordered consumption stays on the main thread. Per-file partials cross threads as compact JSON bytes (jaq Vals are Rc-based, not Send); compiled filters are cached once per pool thread.
• Slurp plans parallelize the parse/wrap phase only — the merged jaq array must be built on one thread.
• emscripten (playground) keeps the fully sequential engine; the worker-only helpers are cfg-gated off that target.

Measured (medians of 5, M-series)

Real cache — 97 docs / 114 MB / 46k steps:

Query	before	after	speedup
length	898 ms	292 ms	3.1×
map(select(.dead_end)) | length	957 ms	309 ms	3.1×
map(select(.step.actor | startswith("agent:"))) | length	976 ms	310 ms	3.1×
.[] | .step.id (stream)	926 ms	303 ms	3.1×
--source claude 'length'	548 ms	200 ms	2.7×
map(.step.actor) | unique (slurp)	1021 ms	856 ms	1.2×
sort_by(.step.timestamp) | .[:5] | map(.step.id) (slurp)	1037 ms	878 ms	1.2×

Even synthetic cache — 60 docs / 56 MB / 132k steps (no skew):

Query	before	after	speedup
length	966 ms	206 ms	4.7×
map(select(.dead_end)) | length	1042 ms	219 ms	4.8×
.[] | select(.dead_end) | .step.id	1063 ms	221 ms	4.8×
map(.step.actor) | unique (slurp)	1236 ms	1092 ms	1.1×

Re-measured 2026-07-23: the original real-cache numbers were taken on a
220 MB cache whose single 110 MB document (48% of the bytes) floored the
parallel wall at 2.0–2.1×; that document has since been deleted. The
synthetic table is unchanged.

Output unchanged

Every benchmark run diffed old-vs-new stdout: byte-identical in all cases (both caches). Unit tests additionally pin the parallel building blocks (render_file, partials_file + unpack_partials + finish_decompose) byte-for-byte against the sequential engine, including compact/pretty/raw modes and the zero-file Decompose rule; the sequential engine remains pinned against slurp by the existing tests.

Verification

• scripts/quality_gates.sh: 7/7 pass
• cargo check --target wasm32-unknown-emscripten -p path-cli: clean (no new warnings; worker-only helpers cfg-gated)
• path-cli: 481 tests pass

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is disabled.}

[github-actions]

🔍 Preview deployed: https://4291e5f8.toolpath.pages.dev