Beagle SCM

BE-055: .be/ dir for flat worktrees — kill the dir⇒own-store proxy in anchor classification

Today a flat (Edwardian) worktree carries .be as a FILE (the wtlog, row-0 referencing the store elsewhere) and a .be/ DIR means "own store here" — ANNO-001 wants a flat wt to carry a .be/ DIR too (wtlog + LSM index runs, NO shards), so every site using .be's kind as the store-vs-flat proxy misclassifies. Full path inventories from the code audit: js-audit (be/ + jab bindings) and c-audit (beagle/ native). Two load-bearing classifiers own the decision — resolve_hash.js anchors() and its native twin dog/HOME.c home_anchor_resolve — with ~40 downstream sites inheriting it.

Input

Context

• JS core break: resolve_hash.js anchors() (61–70) takes kind=="dir" as "own store in .be/" and never consults a row-0 redirect; discover.js resolveAnchor (101–131) dir branch sets storePath = wt — a flat .be/ dir gets store==wt, wrong.
• Native twin break: dog/HOME.c home_anchor_resolve (637–708) dir ⇒ h->root = h->wt; home_walk_up (938–964) anchors a .be/+wtlog dir as primary-store-here. Same wrong root.
• Store-shape consumers scan .be/ expecting shards: store.shardDir readdirs subdirs as project shards; home_be_subdirs/home_single_shard classify shield vs store; serve.serveStore, submount.resolveLocalStore, get.resolveLocalSource, fetchleg.isWtPath all fork on reg⇒wt / dir⇒store.
• Submount detection is FILE-only: sniff/SUBS.c SNIFFSubIsMount (1120–47) and sniff/AT.c dirty-scan prune (932–40) mark a mount by a .be FILE — a flat-dir sub is not pruned/listed; subs_path_is_be_store (332–37) calls any .be-dir a store; beagle/SUBS.c unmount FILEUnLinks the anchor (510–24) — cannot remove a dir.
• The work view lists only .be-FILE worktrees (work.js listWork 72–86); keeper/RECV.c (487–95) gates the colocated be get ? advance on .be/wtlog being a regular file — a flat .be/ dir would pass it with a REMOTE store.
• ~12 row-0 store-reference readers (JS + 2 native) parse the .be FILE's first row; for the dir shape that row lives in .be/wtlog — full list in both audits.
• Precedent: SUBS-049 already half-recognizes a .be/ dir with a wtlog member (classify.js 93–97, work.js isMount, home_walk_up 946–55) — but those branches still anchor store==wt, which is exactly the broken half.
• graf/keeper/spot .be/ references are STORE-side (shard dirs, .graf.idx/.keeper runs via HOMEBeDir) — unaffected except through HOME's root resolution. graf/DAG.h 16–21 layout comment is stale (branch dirs retired, see GRAF.c 42/164).
• dog/git/IGNO.c igno_is_meta skips .be by NAME (kind- agnostic — survives); its ..be.idx sidecar rule goes dead for dir-shaped flat wts (the index moves inside .be/).

Goals

• A flat worktree may carry .be/ as a DIR (wtlog + index runs, no shards) and classify as the Edwardian anchor EVERYWHERE: resolvers, store openers, work view, submounts, wire serve, native and JS in lockstep. Store root comes from the row-0 reference in .be/wtlog.
• Legacy .be FILE worktrees keep working unchanged; own-store .be/ dirs keep working unchanged.

Constraints

• ALL resolution stays in resolve_hash.js + the ONE native twin (HOME.c); consumers must not grow their own kind tests.
• Repro tests FIRST: the three shapes (file / flat dir / store dir) driven through both resolvers before any consumer changes.
• Smallest possible diffs at ~40 sites; classification core lands first, consumers follow; end-to-end verify on real clones.

WIP

Design decisions

• The discriminator for a .be/ dir needs a RULING — candidates: (a) shard presence: any non-dotted SUBDIR in .be/ ⇒ own store, none ⇒ flat; (b) the .be/wtlog row-0 URI: references elsewhere ⇒ flat, self/absent ⇒ own store; (c) an explicit marker member. (a) is cheapest and matches home_be_subdirs; (b) is the semantic truth but taxes every stat with a row read.
• Submount detection moves from ".be is a FILE" to ".be exists and anchors" (presence + classification), and unmount handles both shapes (unlink vs recursive rm).

TODOs

• 1. Ruling: pick the flat-vs-store discriminator for .be/ dirs (subdir scan vs wtlog row-0 vs marker member).
• 2. Repro tests: three-shape table through resolve_hash.js, discover.resolveAnchor, and dog/test/HOME.c's kinds table (extend K_DIR/K_FILE with K_FLATDIR) — failing first.
• 3. JS classification core: anchors() + resolveAnchor learn the flat-dir branch; store root from .be/wtlog row 0; the ~12 row-0 readers routed through the one reader.
• 4. JS store-shape consumers: store.shardDir, serve.serveStore, submount.resolveLocalStore, get.resolveLocalSource, fetchleg.isWtPath stop kind- forking and ask the resolver.
• 5. JS wt surfaces: work.js listWork/isMount accept flat-dir wts; submount.mount/rollback and get's sub unmount handle dir anchors (recursive rm, not unlink).
• 6. Native core: home_anchor_resolve + home_walk_up flat- dir branch (root from wtlog row 0); RECV.c colocated gate re-derives the store instead of trusting wtlog-is-reg.
• 7. Native consumers: BE.cli.c source/cwd classify (788–849, 2541–56, 3378–88), sniff/AT.c + sniff/SUBS.c submount checks, beagle/SUBS.c unmount, subs_path_is_be_store.
• 8. Flip/extend the pinning tests listed in both audits (be/ test fixtures, sniff/test SNIFF.c 1014–33, keeper tests).
• 9. Docs: DOG.h 100–144 layout doc, stale graf/DAG.h 16–21 comment, wiki Worktree/Store pages.

Blockers and bummers

• TODO-1 ruling gates everything; (a)-vs-(b) changes the hot-path cost of every repo discovery.
• Two implementations must move in lockstep or clones disagree about where the store is — the worst failure shape there is.
• ANNO-001 TODO-2 is blocked on this ticket; the anno lane lands on top of the reclassified anchor.

Outcome

• Open. Filed 2026-07-24. Path inventories: js-audit, c-audit (Opus code sweep over be/, jab/, beagle/).