Beagle SCM

LITE-033 DONE: lite lindex — backlink suspects in the index lane

A separate indexing round answering "who links to this page" with a SUSPECT list, not proof: key dst_hl:40|0:20|7, val src path_hl:40|0:20|vnib:4 in the one .lite.idx lane (INDEXES). Only TIP blobs are scanned, lazily — new blobs only. Stale rows are legal: a removed link leaves a false suspect the caller kills by opening the file, so the lane's never-delete, idempotent contract holds untouched. Links are the Link grammar; the recognizer is LITE-026's hook scanner, reused, never a second regex.

Input

Context

Backlinks are the one wiki query the lane cannot answer today: REV/B2P rows say what a path IS, nothing says who POINTS at it. The Link grammar already fuses in-prose refs (main.js, abc/Makefile:20, LITE-029, DOG-036/LINK.rl) in index/hook.js, and index/resolve.js already descends FSEG for partial paths — indexing is a walk over tip blobs with both in hand.

• Ruling (2026-08-15): suspects only — precision comes from opening the suspect file, never from row deletion or tip-filtering hops.
• Ruling: lazy/incremental, NEW blobs only; a full re-walk must write nothing on a second run.

Goals

lite lindex brings the backlink rows up to the current tip; lite lindex <target> prints the suspects for one target.

• New record kind LINK (7): key dst_hl:40|0:20|7, val src path_hl:40|0:20|vnib:4; one row per (dst, src) pair, re-puts idempotent.
• dst_hl is minted from the target TEXT: repo-relative path text for file links, the bare ticket code (LITE-029) for ticket links — so a ticket's backlinks survive the thin↔fat layout move.
• Incremental mark: a MARK-style row under the reserved ref hlOfText("lindex") holding the last-processed commit; a run diffs mark..tip and scans only the changed paths' new tip blobs, then moves the mark LAST.
• Query lite lindex <target> hashes the target text, one prefix scan, and prints each suspect as a path (via FSEG/tree descent or a kept sidecar — worker's call, but the answer is TEXT); no rows for a target prints nothing.
• Update lite/INDEXES.mkd: the LINK kind, the suspects contract, the mark.

Constraints

• Ruling (2026-08-15): link recognition = the STANDARD tokenizer, which fuses a link into one F token (tok.parse, the Link grammar) — lindex reads a blob's F tokens the way index/hook.js does, and never re-scans raw bytes; partial-path resolution = index/resolve.js FSEG descent. NO link regex, NO second resolver.
• Unresolvable or ambiguous refs are SKIPPED, hook-style — the indexer never guesses; a self-link mints no row.
• Only prose-bearing blobs need scanning; binary blobs skip. Anchors are dropped: the row names files, not places.
• Verb shape follows lite's: thin main.js case over an index/lindex.js body; errors in plain words to fd 2.
• Tests: test/lindex/run.sh under dash, fixtures in $HOME/tmp; repro-first — the no-backlink-answer gap, then the verb. ZERO quickjab edits.

WIP

Design decisions

• Suspects-only keeps the lane append-only: correctness lives at the caller, the index only narrows the grep.
• vnib stays 0 (reserved); the val's spare 20 bits stay 0.
• Tip moves re-scan only changed paths' new blobs; rewritten history needs no special case — the mark row simply jumps.

TODOs

• repro: a fixture repo where "who links to X" has no answer today
• factor the tokenizer F-token reader out of hook.js for reuse
• index/lindex.js: the mark..tip walk, LINK rows, mark write LAST
• dst minting: path text vs ticket code; FSEG descent for partials
• query form: prefix scan, suspects printed as TEXT paths
• INDEXES.mkd: LINK kind, suspects contract, lindex mark
• test/lindex/run.sh: add/remove-link, incremental no-op, ticket dst

Blockers and bummers

• Path TEXT back from a 40-bit path_hl: no sidecar — the query descends the tip tree once and prints the paths whose hash the rows named, so INDEXES.mkd's "NO path text" gap stays literally true.
• Deviation, gritzko's call: dst mints by resolve-at-tip — one hit keys the path text, many hits skip, ZERO hits key the ref's own text verbatim. Ticket codes are exactly the zero-hit case (no code-shape predicate in lindex), but an unresolvable FILE ref then also mints a verbatim row instead of being skipped as the ticket ruled.
• test/lindex needs a quickjab runtime ($LITEJAB): only the DOG-034 lexer fuses dir/file.c into one F token; under plain jab the suite fails.
• Pre-existing at base 6b1e478c, not this ticket: blob/cat/list/tree run.sh plant $WORK/jsrc too late; the main tree already carries the fix.

Outcome

Staged in work/LITE-033, 8 files, nothing posted. Orchestrator verification, independent of the worker's harness:

• test/lindex/run.sh PASS, 31 checks (15 shell + 16 rows.js) under quickjab; worker ran it red-first (15/15) before the verb existed.
• Touched suites re-run clean: hook 14+cycle 7, index 129; worker tallied the full suite against a pristine base copy — identical, no regressions.
• Real-path smoke on a scratch git repo: scan mints 3 rows for 2 links, lindex main.js and lindex LITE-033 both answer doc.mkd, a bare rerun says up to date and the lane files stay byte-identical.
• Files put: INDEXES.mkd, README.mkd, index/hook.js, index/index.js, index/lindex.js, main.js, test/lindex/rows.js, test/lindex/run.sh.
• Suggested commit message: LITE-033: lite lindex — backlink suspects in the index lane