[thekevinbot]

Refs #1483

Per-table on_file hooks now start on the first path the startup walk matches and are fed paths as it finds them, instead of after the walk ends. Hook-spawn latency and path copying overlap the walk. Output is unchanged; row order across invocations was already unspecified.

Measurements (65k-note vault, GROUP BY status, release binary, A/B interleaved vs PR A head)

• Unrestricted CPUs (load ~7-13): min 82.4 -> 79.6ms, median 85.7 -> 88.9ms (within noise).
• taskset -c 8-15 (load ~16): median 103.5 -> 87.2ms.
• taskset -c 8-11: median 150 -> 132ms.
• strace: first awk spawn at 53ms -> 23ms after process start.
• Path-table query (no hook): 18.6 -> 19.1ms (noise). jsonl 692 -> 683ms (noise).
• Harness ratio vs fd+xargs -P stays ~2.5x, so the bar is not flipped: the remaining cost is dirsql's own walk (~30ms), SQLite ingest and process start, not the spawn.

Decisions

• The hook thread peeks its first path and does not call the hook for an empty feed, keeping "a table with no files spawns nothing".
• Persist builds keep the old after-scan path (cache reconciliation needs the full scan first).
• Bar for Speed: frontmatter parse is 3.2x the fastest native (find + xargs -P) #1483 left as is: local run does not cross 1.1x.

Red/green

Red run 38190251372 (Unit tests job): prepare_starts_a_per_table_hook_while_the_walk_is_still_running failed at assert!(hook_saw_it_mid_walk.get()), the asserted reason.

E2E Verification

• uv run python -m pytest tests/e2e/jsonl_parse_speed_test.py tests/e2e/frontmatter_parse_speed_test.py tests/e2e/on_file_query_test.py tests/e2e/query_contract_test.py -x -q: 14 passed (python receipt).
• pnpm test:e2e: passed (ts receipt).
• cargo test -p dirsql --lib: 1152 passed. Gated preflight: 0 failing.