[thekevinbot]

Refs #1474

Path-table rows are now built on the walk worker threads, so building them overlaps the walk instead of running serially afterwards. Each row is also smaller: 80 bytes instead of 104. The path is a Box<str>, the scanned path a Box<Path>, and the basename/dir/ext spans are niche-packed NonZeroU32. Output is byte-identical (same md5 over all 341k rows).

Timings

SELECT path FROM './**/*.md' on the 400k-file glob_speed_recursive tree (341k matches). Runs are paired and interleaved, 40 each, medians shown. Other agents were running, load 3 to 8.5.

binary	median	RSS
main (462b3c6)	72.1 ms	107 MB
this PR	61.6 ms	77 MB
fd --no-ignore -t f -g '*.md'	40.7 ms	20 MB

That takes the gap from 1.77x fd to 1.51x. Phase breakdown (stderr stamps):

phase	main	this PR
walk	~17 ms	~21 ms
row build	~13 ms	0 ms (inside the walk)
SQL loop + JSON	~28 ms	~27 ms
exit	~6.5 ms	~4.4 ms

I measured both parts separately. Rows-in-walk alone gave 71.5 to 60.8 ms (hyperfine). The row shrink gave a further 69 to 63 ms against rows-in-walk (hyperfine), and fewer page faults (22.9k to 19.4k).

Remaining gap (not in this PR):

• SQL loop (~27 ms). Profiling shows fill_cell at 28% self time (from_utf8_lossy plus a copy into Value::Text), sqlite3Step at 20%, value_ref at 15% and serde string escaping at 13%. Writing JSON straight from ValueRef text would skip the copy and the validation.
• Exit (~4-5 ms). This scales with RSS.

E2E Verification

• Python, targeted at the changed path table and walk: testing-conventions e2e attest 'uv run python -m pytest tests/e2e/glob_speed_recursive_test.py tests/e2e/glob_speed_one_level_test.py tests/e2e/glob_speed_depth_bound_test.py tests/e2e/glob_speed_gitignore_test.py tests/e2e/glob_speed_dotfiles_test.py tests/e2e/glob_speed_matchers_test.py tests/e2e/glob_speed_node_modules_test.py tests/e2e/glob_speed_symlink_test.py tests/e2e/stat_speed_test.py tests/e2e/query_contract_test.py -q'. Result: 26 passed. The module was built with uv run maturin develop --release and cargo build --release -p dirsql --features cli.
• TypeScript: just e2e-attest-ts. Result: 18 passed.
• cargo test --workspace --features cli: pass.
• cargo bench -p dirsql: ran clean.

Manual run: dirsql query with path, basename, dir, ext, size and content, over root-relative, ../ and absolute globs. Every column reads correctly, and the full ./**/*.md output matches main byte for byte.

Changelog / Migrations

• packages/rust/changelog.d/2026-10-13-path-rows-built-on-walk-threads.md
• packages/rust/migrations.d/2026-10-13-path-rows-built-on-walk-threads.md: not breaking, all subsections _None._