Beagle SCM

SUBS-021: SHA leaks into the <subproj> slot on a re-posted sub dot-branch

When a submodule POST targets an ALREADY-synthetic / detached-sha target, the synthetic dot-branch is mis-assembled: the commit SHA leaks into the <subproj> slot AND into the #fragment, producing a malformed ref ?<sha40>/.dogs#<sha40>/.dogs instead of ?/<sub>/.<parent>#<sha>. The correctly-formed branch ?/abc/.dogs (with the /) is fine — the earlier "missing /" reading was a ripgrep-rendering misread (2e=.); the real bug is the SHA in the subproject slot. Method: Issues.

Input

Context

• CORRECTION (2026-06-16 scan): the synthetic dot-branch is built CORRECTLY as

  ?/<sub>/.<parent> — beagle/BE.cli.c:3069-3080 (bepost_synth_child_uri) emits / then . (lines 3073-3074); the abc sub's good commits sit on ?/abc/.dogs in both .be/abc/refs and abc/.be (hexdump 3f 2f 61 62 63 2f 2e 64 6f 67 73). Spec Submodules:60-68 agrees. The original "missing /" premise is INVALID.

• REAL defect: one stray malformed row — .be/abc/refs 266168WPCD (mirror

  abc/.be:26): post ?<sha40>/.dogs#<sha40>/.dogs — the <subproj> segment is the 40-hex SHA, not the title abc, AND the #fragment carries <sha40>/.dogs instead of a bare sha.

• Suspected cause: a re-POST whose target is already synthetic/detached feeds a

  detached-sha "project" into the subproj slot. Leads: bepost_wt_project (BE.cli.c:2954-2990), the lq/DOGQueryStripProject arm (:3010-3025), the synth_view/msg_view assembly (:3159-3201).

• This stray row is the likely source of the abc sub's odd be head "2 ahead" /

  POSTDET symptoms hit while landing ABC-003 — NOT a branch-format bug.

Goals

• The <subproj> slot is ALWAYS the project title (never a SHA) and the

  #fragment a bare sha — on first post AND on a re-post of an already synthetic/detached target.

• A repro re-posting a sub with an already-synthetic/detached target, asserting

  the recorded branch is ?/<sub>/.<parent> with a bare-sha fragment.

Constraints

• Behaviour-preserving for correctly-formed first posts. beagle//sniff store

  changes gate for review. Add the repro to the SUBS table-driven tests.

• The single stray malformed ref in the REAL store (~/beagle/.be/abc/refs

  266168WPCD + abc/.be) is a SEPARATE maintainer cleanup — the worker must NOT touch ~/beagle or its store; the good commits are independently reachable on ?/abc/.dogs, so dropping that row orphans nothing.

WIP

Design decisions

• ROOT CAUSE FOUND: NOT the POST composer (bepost_synth_child_uri is correct —

  always emits ?/<sub>/.<parent>#<bare-sha>). The leak is GET-side: sniff/GET.c::SNIFFGetURI raw-hex fallback (~:2213) accepted ANY non-empty u->query as a checkout target with no hex check. A be get '?<sha40>/.<par>' missed REFSResolve, fell through, and the fallback wrote the query string VERBATIM into BOTH the recorded query AND fragment → ?<sha40>/.par#<sha40>/.par (it "succeeded" only because the 40-hex prefix matched the pin via prefix lookup).

• FIX: gate the fallback on DOGIsHashlet(u->query) (6..40 hex). A query

  with path junk (?<sha40>/.par) is not a hashlet → clean SNIFFFAIL; clean sha-PREFIX gets (?abc1234, README:64) still work. (NB: DOGIsFullSha — used in the first pass — was WRONG; it requires 40/64 hex and would reject every sha-prefix get.)

TODOs

• Repro-first: test/post/41-sub-synth-no-sha-in-subproj/ (hermetic file://)

  — RED pre-fix (sha-in-subproj get accepted rc 0), GREEN post-fix (refused).

• Root-caused: GET-side fallback, not the POST arm (above).
• Fix: sniff/GET.c DOGIsFullSha gate; behaviour-preserving for real ?<40hex>.
• SUBS/post table-driven repro added; 299/299.
• Landed 8c69c16b (be put+be post --nosub; 299/299 before & after).
• (maintainer) drop the stray real-store rows; NOT from the wt.

Blockers and bummers

• Separate, OUT OF SCOPE: be get '?/<sub>/.<parent>' (well-formed synthetic

  re-attach) currently also fails — a reconcile-to-synthetic-branch issue; flag as a follow-up, not fixed here.

• ABC-003 is NOT blocked by this (dot-branch well-formed; wt already attached).

Outcome

• LANDED 8c69c16b: sniff/GET.c DOGIsHashlet gate (corrected from

  DOGIsFullSha) + test/post/41-sub-synth-no-sha-in-subproj (malformed ?<sha>/.par refused AND a clean sha-prefix get still succeeds); 299/299 before & after; canonical untouched. Follow-up still OPEN: be get '?/<sub>/.<parent>' well-formed synthetic re-attach also fails (reconcile-to-synthetic-branch issue).