Beagle SCM

STATUS-002: be status ignores only the sub's own .gitignore — needs hierarchical (cwd→$HOME) compounding

be status:<sub>/ floods with unk for paths the repo-root .gitignore should exclude (build-*/, *.o, CMakeFiles/, …): the relayed sub-scan IGNOLoads only the SUBMODULE's own .gitignore, which abc lacks, so the parent's patterns never reach the sub's paths. Bare be collapses a sub to one summary row (no flood) — that's the "intermittent". Fix: hierarchical .gitignore resolution — load every .gitignore from the classified path's directory UP to $HOME (or / if outside $HOME) and COMPOUND them, crossing submodule boundaries, the way nested git ignores stack. NO hardcoded directory names. See Sniff.

Input

Context

• Repro (deterministic; same on PATH be and a fresh build — NOT skew): build

  the abc sub (makes abc/build-debug/{CMakeCache.txt,*.o,…}), then be status:abc/ → ~15 unk abc/build-debug/... rows.

• be status:<sub>/ relays into the sub's worktree; its SNIFFOpen/IGNOLoad

  reads only the SUB's own .gitignore (absent for abc), so the repo-root .gitignore never governs the sub's paths.

• Bare be collapses the sub to one summary row → no flood (the "intermittent").
• POST-020 (8c7f5f3a) REFUTED: sniff/SNIFF.c/CLASS.c/dog/git/IGNO.c/

  ULOG.c are byte-identical to parent d8fec51a; the bug is pre-existing / structural (since submodule projector recursion), not a regression.

• The repo-root .gitignore already lists every flooded pattern — it just

  isn't consulted from inside the sub.

Goals

• Hierarchical .gitignore: classifying a path loads every .gitignore from

  its directory up to $HOME (or / if not under $HOME) and COMPOUNDS them (git precedence: nearer/deeper rules add or override; ! negation honored).

• The chain CROSSES submodule boundaries, so a parent repo's .gitignore

  governs nested sub paths — this is what fixes the abc/build-debug flood.

• No hardcoded dir names; build/Corpus/etc. are excluded ONLY because the

  real .gitignores in the chain list them.

Constraints

• ROLL BACK the prior stop-gap (unconditional build*//Corpus/ ignore in

  IGNOMatch): Corpus is not special — restore pattern-only matching.

• Honor git pattern semantics (anchoring, dir-only trailing /, ! negation,

  precedence) across the compounded chain.

• Terminate the walk at $HOME (or /); cache the chain per directory so the

  status walk doesn't reload .gitignores for every leaf.

WIP

Design decisions

• REJECTED (the prior worker's landed-in-tree fix, to be rolled back):

  igno_is_build/Corpus special-case in IGNOMatch — a hardcoded dir blacklist, not ignore resolution.

• CHOSEN: an .gitignore STACK loaded leaf-dir → $HOME//; IGNOMatch

  consults it nearest-first with git precedence; built once per directory (cached), not per leaf. Localize in dog/git/IGNO + the sniff/SNIFF open/scan that drives it (incl. the be status:<sub>/ relay).

TODOs

• Roll back the prior fix: drop igno_is_build*/Corpus from dog/git/IGNO.c,

  revert the test that pinned it; IGNOMatch pattern-only again.

• Implement the chain: load+stack .gitignore from the path's dir up to

  $HOME//; compound with git precedence; cache per directory.

• Wire into the status walk so sub subtrees inherit the parent chain.
• Repro-first test (hermetic): a sub subtree with a PARENT-ONLY .gitignore

  → build artifacts ignored; plus a !-negation / precedence case. ctest green.

Blockers and bummers

• Prior worker flagged the STATUS-002 worktree / at 100% disk (built via a

  tmpfs symlink). Watch disk on the rerun.

Outcome

• (open; prior build/Corpus stop-gap rejected, to be rolled back.)