DIS-025: be patch squash uses the target's PARENT as merge base — silently drops commits when cur is behind the target

Now
DONE

HIGH — silent data loss. STAGE 1 LANDED (a31c7fbd); data loss FIXED. Remaining: the KEEPResolveURI funnel consolidation (Stage 2/3 below) — a separate session. Original report follows. be patch ?<target> computes the squash merge base as LCA(target's parent-branch tip, target), which for a target commit resolves to the target's immediate parent. That is correct only when cur is a descendant of that fork (the intended "absorb a feature branch's stack into its parent" case). When cur is an ancestor of the fork — i.e. cur is behind the target (reconciling a stale clone, or patching a trunk commit that is ahead of cur) — every commit in (cur .. fork] is NOT in cur, yet the per-file classifier treats it as already-in-base and silently drops it. be patch exits 0, failed=0. (be get's reconcile hits a related but distinct symptom on the same inputs — sniff: merge failed … (graf err) — see "Related" below.)

Repro (be patch only — confirmed, instrumented)

Trunk: eb1945fb (base) → … → 50722656 (SUBS-008, edits beagle/SUBS.c) → 50724d9a (SUBS-009, target). cur is a branch at eb1945fb.

  1. be put ?ours#eb1945fb; be get ?ours (cur = eb1945fb).
  2. be patch ?50724d9a.
  3. Result: exit 0; beagle/SUBS.c reported mod, on-disk == cur's bytes; 008's FILErmrf is gone. Only 50724d9a's own last-commit delta (SUBS-009) was absorbed; SUBS-002/006/007/008/010 (in eb1945fb..50722656) were dropped.

Root cause (confirmed by instrumenting fork_sha)

sniff/PATCH.c ~L1533-1543:

resolve_parent_tip(&parent_tip, reporoot, target_query);   // → target's parent branch
GRAFLca(&fork_sha, &parent_tip, &thr_sha);                  // → 50722656 (target's PARENT)
if (sha1empty(&fork_sha)) GRAFLca(&fork_sha, &our_sha, &thr_sha);  // fallback only if EMPTY

Instrumented output for the repro:

DBGFORK our=eb1945fb  thr=50724d9a  fork=50722656

The fork is 50722656 (target's parent), not the true base eb1945fb (cur). The per-file 3-way classifier patch_walk (sniff/PATCH.c:591) then sees, for SUBS.c, fork-blob == theirs-blob (both 949938ea, the post-008 version) while ours differs → branch !o_eq_l && t_eq_l → mod ("only ours changed; theirs didn't touch"), keeping cur's stale bytes. The same fork explains every file in the run: PUT.c (fork==ours, theirs≠) → take-theirs; CMakeLists (all three differ) → merged.

The code's assumption (its own comment): the parent-tip LCA "excludes ancestor commits already in cur's history." That holds only when cur ⊇ [fork..parent_tip]. It does not validate that fork is an ancestor of cur. When cur is behind the fork, the "excluded" commits are genuinely absent from cur and are lost.

Why blobs / earlier probes did NOT reproduce it (history-dependent)

Expected

The 3-way merge base for absorbing theirs into cur must be a common ancestor of cur and theirs: LCA(our_sha, thr_sha). Anything in (base..theirs] that is absent from cur must be absorbed (or surfaced as a conflict), never silently dropped with exit 0.

Deeper root cause (refined; reconciled against URI.mkd)

The fork came out as the target's parent because be patch ?<sha> resolves the bare sha to the canonical detached query /<project>//<sha> — an empty branch segment (//) between project and pin. Per the authoritative spec this form is correct, not malformed:

> URI.mkd §"Resolution boundary": A detached or branchless target (tag > checkout, ?null, bare sha) has an empty branch slot: a double slash > /<project>//<full-hash>.

So be emits exactly the right canonical string. The bug is downstream parsing that refuses it. dog/DOG.h DOGCanonQueryParse (~L494-499) explicitly rejects the empty-branch form:

//  zero-segment (e.g. "/proj//sha") is rejected as malformed.
if (branch_start >= branch_end)      return NO;

Because the canonical parser returns NO, DOGRefSplitPin drops to its manual fallback, which hands the whole /<project>/ back as the branch and the sha as the pin. The cherry-locate detector (sniff/PATCH.c ~L1444) then misreads this detached SQUASH target as a ?<branch>/<sha> cherry-locate and promotes it (cherry = YES); cherry-pick sets fork = parent(sha) — hence the dropped intermediate commits. Confirmed: target_query=[/30-cur-behind-target//<sha>], shape=SQUASH cherry=1. Both the short hashlet and a full 40-hex bare sha trip it; an explicit ?<branch>/<sha> (single slashes, non-empty branch) is unaffected.

The earlier "stop emitting //" idea was backwards: the spec requires the // form; the emitter is correct and the parser is wrong.

Fix plan: one canonical resolver, empty-branch contract first-class

Reference resolution is currently scattered across three layers that each re-derive structure from the URI string — dog (DOGNormalizeArg, DOGCanonQueryParse, DOGRefSplitPin, DOGRefIsBranch), keeper/RESOLVE (KEEPResolveRef → bare sha, kind discarded), and per-verb ad-hoc splitting (PATCH's // cherry detector). DIS-025 is the symptom of that duplication: the kind (branch vs tag vs detached-hash) is computed, thrown away, then guessed again downstream — wrongly.

Target architecture — a single keeper-owned funnel. Resolution moves behind one entry point that turns raw user URI text into fully-canonical absolute URI text, reading cwd + struct home for context:

//  keeper-level: the only layer that sees home + REFS + pack index.
ok64 KEEPResolveURI(home *h, u8s abs_uri, u8cs rel_uri);
//  three contextual arms resolved at once:
//    path  — cwd-relative (./f, ../d, bare f)  → project-root-relative
//    ref   — ?., ?.., ?back, bare, tag, hashlet, empty-branch detached
//            → /<project>/<branch>/<40-hex>   (// when branch empty)
//    auth  — //alias → full transport URL      (folds in REFSResolve grep)
ok64    REFSResolveURI(home *h, u8s abs_ref, u8cs rel_ref);  // ref-only sub-step
refkind REFSQueryKind(u8csc canon);   // cheap kind probe, no alloc/lookup

home already carries everything the context arms need (project, cur_branch, cur_sha). Text-in/text-out makes the funnel a table-testable string→string canonicalizer; bareword promotion (verb-specific) stays upstream in be/dog, so rel_uri always arrives marker-bearing and the funnel is verb-agnostic. The embedded pin doubles as the concurrency guard, retiring the separate --at #sha. KEEPResolveRef/KEEPResolveHex shrink to wrappers; REFSResolve's alias grep becomes the //auth arm; the DIS-025 trio (DOGCanonQueryParse empty-branch reject + DOGRefSplitPin fallback + PATCH's // cherry guard) all collapse into one canonical producer + one parser that treats /<project>//<sha> as first-class detached.

Staged implementation.

  1. Empty-branch contract (foundational, fixes DIS-025). — LANDED. Commits b4c6397f (test hermeticity, see below) + a31c7fbd (the fix) on beagle trunk. Two changes: Note vs the earlier draft: no REFSQueryKind and no fork-ancestor guard were needed — the peel emits a full 40-hex, so the existing resolve_parent_tip 40-hex bail already yields the correct base. Verified: repro test/patch/30-cur-behind-target RED (a.c kept at v0, C1 dropped) → GREEN; all 21 be-patch-* green; full debug suite 232/232.
  2. KEEPResolveURI + REFSResolveURI + REFSQueryKind with a hermetic table test (path arm, ref arm, auth arm; empty-branch detached among the cases). Wire PATCH's target resolution through it (replaces the canonic-peel + DOGRefSplitPin cherry detector at sniff/PATCH.c ~L1405/L1435).
  3. Caller migration. Route GET / POST / be dispatch through KEEPResolveURI; retire DOGRefSplitPin and the scattered re-derivation.

Stage 2 starting points (for a cold session)

The original reconcile used be get (which bases on LCA(our, theirs) = eb1945fb, the correct base) and still failed — sniff: merge failed for beagle/SUBS.c (graf err), leaving ours untouched. That is a distinct failure in GRAFMergeWtFileTunable's weave apply on a correct base (it errors loudly rather than dropping silently); track separately if it recurs. The landed 7d37f90f is correct only because that reconcile was completed out-of-band (git merge-file) and verified token-for-token.

Severity / blast radius

Any be patch ?<target> where cur is behind the target's parent-branch fork — the exact shape of the serialized fix-and-land loop (a clone cut at an older trunk, patching/reconciling a newer trunk commit). Same data-loss class as SUBS-001; worse because it is silent (exit 0).

Outcome

RESOLVED — delisted in the 2026-06-25 triage.

❧ RSS feed ❧ Beagle SCM ❧ Beagle for git ❧ JAB runtime ❧ QuickJAB runtime ❧ BlueSky ❧