SUBS-027b: submodule recursion semantics + record-writing + the deferred classifier axes (follow-on to SUBS-027)

SUBS-027's first slice landed the SUBSDirty classifier + be status Advanced-detection. This ticket is the deferred half: actually WRITE the R1/R2 records, compute the Edited and coord-mismatch axes, enforce the recursive get/post return-code semantics, surface the pin axis in be head, and fill the get/post test gaps. Serialized AFTER SUBS-026 + PATCH-001 (same gitlink-bump path). See Submodules, SUBS-027.

Input

Context

The landed slice classifies + flags be status, but three pieces are placeholders and the verb semantics are unenforced:

R1/R2 are not written by any code. SUBSDirty reads R1 as the live tree gitlink and probes R2 (usually ABSENT — your beagle case). The parent-pin reflog mirror and the ?.<parent> synth keep-alive must actually be emitted on post.
Edited (bit 0) is declared but a no-op — the recursive cross-store file-scan ≠ R4.base comparison is unwired, so today only Advanced drives status.
coord-mismatch unwired — the R2.<parent/branch> vs R3.branch check (record-health axis) is not computed.
Recursion semantics unenforced — the get reapply-onto-pin and the post OK/POSTNONE/error contract live only in the Submodules spec; sniff/POST.c + beagle/SUBS.c don't implement them. These files COLLIDE with in-flight SUBS-026 and PATCH-001.
be head can't see the pin axis — graf/LOG.c computes changed as committed-tree-vs-branch-tip; it structurally cannot see wt-vs-pin (the tree still pins the old sha). Needs a sniff→graf path.
Test gaps (audit of test/get,post,patch,status): Advanced/behind/diverged + all record-health are untested at GET/POST level (only patch/28,/35 cover advanced/behind); post/26+post/27 are WILL_FAIL stubs already encoding the post return-code contract, gated on "SUBS-001"; get/14 gates on an unimplemented be get --dry-run.

Goals

Write R1 (parent-pin reflog mirror) and R2 (synth dot-branch) on post; the classifier reads real records, the tree stays authoritative for R1.
Compute the Edited (recursive) and coord-mismatch axes so all three SUBSDirty axes are real.
Enforce the get (reapply onto pin) and post (OK→read tip + stage bump / POSTNONE→no bump / any sub error→HARD-ABORT the whole post; base≠pin-on-entry→error unless --force) recursion contract.
Surface Advanced/stale per sub in be head.

Constraints

SERIALIZE after SUBS-026 + PATCH-001 — they rewrite the same gitlink-bump path; concurrent land = non-FF race (work.mkd hazard).
Reuse the landed SUBSDirty (one classifier, every verb); ABC style; ragel for any parsing; no per-verb tree re-parse where R1 suffices.

WIP

Design decisions

—

TODOs

No-quiet-commit-loss (the safety core of the post contract): a sub PROVABLY BEHIND the pin (R4.base a real ancestor of R1) is NEVER auto-bumped — recording the older tip would move the gitlink BACKWARD and drop pinned commits. Landed 719cb085; forward-only via the SUBSDirty Advanced/stale gate in sniff/PUT.c; repro test/post/51-sub-behind-no-backward-bump.
Classifier robustness: SUBSDirty now derives the sub shard from the path basename when the anchor is unbound (be get ?<sha>, the SUBS-028 blank-want), so a behind sub walks real ancestry instead of the old "assume Advanced" guess (which drove the backward bump). Truly-unresolvable shards keep the conservative-Advanced signal (never miss a real bump — verified vs post/23, post/45).
Emit R2 (?/<sub>/.<parent> keep-alive) on post — landed 0c67bea8. The committed case already posts onto the synth branch; the gap was the advanced-CLEAN bump (sub made no commit → never anchored), which now anchors R2 at the new pin via be put ?<synth> in the sub mount, gated on silent (no-commit) so it never perturbs the committed/push recursion (a broader bolt-on regressed 16 push/post tests). Repro test/post/52-sub-advanced-anchors-r2. Closes the R2 arm of SUBS-026.
Emit R1 (parent-pin <sub/path>?<parent/branch>#<hash> mirror in the parent refs) + point the classifier at the records. Lower priority: a cache the classifier doesn't need yet (it reads the pin from the authoritative tree); do it together with a consumer (coord-mismatch).
Compute Edited: recursive cross-store sub file-scan ≠ R4.base, folded up the mount tree (node dirty if it OR any descendant is).
Wire coord-mismatch: R2.<parent/branch> vs R3.branch.
Remaining post-contract pieces: a LOUD error+--force escape (today behind silently keeps the pin + warns when not -q) and refusing a real DIVERGED cousin (today bumped conservatively, since cousin vs undecidable needs object fetch); the get reapply-onto-pin path.
be head: per-sub Advanced/stale mark — landed b9c6b647 ("be head flags sub advanced past pin"; beagle/BE.cli.c:1707-1801 SUBS_DIRTY_ADVANCED → "advanced" mark, tests head/04-sub-dirty + head/05-sub-parent-ahead).
Tests (repro-first, CLAUDE §17): Advanced/behind/diverged + synth-missing/coord-mismatch/unbound at GET and POST; recursive depth combos; UN-GATE post/26+post/27 (the post return-code contract) when it lands; get/14 once --dry-run exists.

Blockers and bummers

Dependency satisfied: SUBS-026 gitlink-bump arm + PATCH-001 both landed; this slice builds on the same sniff/PUT.c gate.

Outcome

No-backward-bump (no quiet commit loss) landed on beagle trunk 719cb085: sniff/PUT.c forward-only gate + sniff/SUBS.c unbound-anchor basename fallback; repro test/post/51-sub-behind-no-backward-bump.
R2 keep-alive emission landed 0c67bea8: an advanced-clean bump now anchors ?/<sub>/.<parent> at the new pin (no synth-missing GC-orphan); repro test/post/52-sub-advanced-anchors-r2; full suite 415/415 (only network CURLtest/POLtest skipped).
The be head pin axis landed b9c6b647 (per-sub advanced mark in beagle/BE.cli.c) — closes the last SUBS-026 residual.
STILL OPEN (audit 2026-07-07): R1 emission (parent-pin cache, defer with a consumer), Edited + coord-mismatch axes (JS twin: DIS-058 D14), the loud error+--force / diverged-refusal, get reapply-onto-pin, and the GET/POST axis test matrix (post/26+27 still gated, get/14 needs --dry-run).