CODE-078: grep at linux scale: the cap counts suspects and the 12-bit slot saturates

Now: OPEN
Sev: MED
Rep: ///bee
Ask: gritzko
Rev: ///bee#8cb69a0
See: BEE-071
See: BEE-066
See: SPOT-002

Input

Context

view/grep.js:64:EJ caps on SUSPECTS, so a needle present in ZERO files is refused: against the real ~~/src/linux lane reftable_stack_destroy (absent from linux) yields 733 candidate files -> "733 files may carry them — narrow the query"; kfree yields 61032 of~~ 95k files in 2.5 s of descent; on git.git strbuf yields 616. index/tridex.js:31:zw: 12 bits of prefix_hl12 saturate — the fills 4096/4096 depth-3 slots in linux, printk/str/buf/EXPORT_SYM all >= 3874 — so the file-level test is noise, and DIR_DEPTHS = 3 gives no dir pruning below drivers/net. Found in the 2026-09-18 index review.

Goals

Confirm candidates under a byte/file budget and cap on CONFIRMED hits; widen the file slot (a u64 lane, or a per-depth trigram block) and/or raise DIR_DEPTHS so deep files still prune.

WIP

Design decisions

The cap moved off the SUSPECTS onto the CONFIRMED hits: the candidates are opened in path order under a per-repo budget of 16384 files and 128 MB, a repo refuses only past 200 CARRIERS ("over 200 files carry them"), and a repo that outruns the budget keeps the old "N files may carry them" with the count it never settled. No lane format changed.
The budget is sized off linux: kfree needs 9452 opens and 70 MB to reach the 201st carrier, so anything tighter would refuse by the suspect count again. OPEN_FILES and OPEN_BYTES in view/grep.js are the knobs.
WIDENING THE FILE SLOT — the options, with the row cost. NONE is implemented: each one changes the entry layout, and that bumps TRI_EXT and TRIDEX_REF together and rebuilds the lane (linux about 9 minutes, BEE-071's own number).
- (a) a u64 lane, tri:18 | depth:2 | hl:24 (up to hl:44). The row goes 4 -> 8 bytes, so linux's 163 MB of .tri.idx becomes ~326 MB; the row COUNT is unchanged. 24 bits are 16.7 M file slots against 95 k files, so the file test stops being noise. Needs a u64 scalar lane in abc.index, a QJAB ticket twinning QJAB-013. The only option that actually cures the saturation.
- (b) a per-depth block in the same u32 lane: files get their own family, tri:18 | hl14:14, dirs keep today's entry. The row stays 4 bytes; it costs one more family and one more row per (file, trigram). 4x the file slots — 16384 against 95 k files — so linux saturates all the same. Cheap, and not enough.
- (c) raise DIR_DEPTHS. Past three levels the depth field needs 3 bits and the u32 has none to spare, so hl12 drops to hl11 and the DIR test halves; it also mints rows for dirs at depths 3+, a format change by itself. It buys descent time (0.8 s of kfree's 8.9 s, warm) and trims the candidates whose dir carries no trigram, but it cannot beat a 12-bit file slot. Free if (a) lands, where depth:3 | hl:24 still fits.
Recommended: (a), with (c) folded into its wider entry.

TODOs

[ ] repro: absent needle answers empty, not refused
[ ] design: wider file slot / deeper dirs

Blockers and bummers

The widening is gritzko's call: (a) needs a u64 scalar lane in abc.index (quickjab), which is outside this repo, so this ticket implements the budget alone.
The budget is not covered by a test: tripping it wants a fixture over 16384 files or 128 MB. test/tridex's leg 5 covers the hit cap and the empty answer instead.