Beagle SCM

JGET-001 UMBRELLA: jab get differential-vs-native bug batch (13 defects)

A differential probe of jab get (the JS be get, be/verbs/get/get.js + be/shared/*.js) against native be get as oracle, retrofitting all ~54 untouched beagle/test/get cases (the JS-096 test-port follow-on), surfaced 13 NEW defects — 3 of them SILENT loss/corruption of committed or uncommitted user work (verified from evidence). The current get.js DOES implement in-worktree forms (inRepoSeed) and recursive submodule mounting (submount.js), so these are real-code bugs, not no-ops — which also makes DIS-058's "JS get never descends a sub" and parts of JS-096's triage stale. Method Issues; one child ticket per defect, grouped by disjoint file ownership.

Input

Context

• Found 2026-06-28/29 by 4 parallel probe workers (native be oracle vs jab get on the same form); evidence under ~/todo/JS096-probe{1,4}, ~/todo/JSGET-cases, /tmp/claude-1000/jabprobe.
• Dedup'd against the known set: JS-097, GET-035 (WONTFIX — atomic, not a bug), GET-036, GET-037, GET-038, DIS-058, JS-075/079. All 13 are NEW.
• 3 CRITICAL repros re-verified from the left-behind worktrees: case 28 jab wt regressed to the old tip; case 58 jab wt = peer's sibling; case 09 jab wrote <<<<dirty||||-feat>>>> into an uncommitted file.

Goals

• Each defect gets a child ticket with a repro-first failing test, a root-cause file:line, and the native oracle behavior it must match.
• CRITICAL data-loss defects (JGET-002/003/004) land FIRST.

Constraints

• Repro-FIRST (CLAUDE.md §17). No hand-rolled URI parsing — fix via the URI binding / parseRemote. Keep changesets file-disjoint where possible.
• Port the matching beagle/test/get assertion into be/test/get as the regression gate (parity vs native, the JS-096 harness model).

WIP

Design decisions

• Severity-ranked children (file:line in each):
  • JGET-002] CRIT no fast-forward/ancestry gate → clobbers committed local

  commits (cases 27/28/58); get.js seed*/fanoutWholeTree, no dag.aheadBehind.

  • JGET-003] CRIT be get ? (trunk switch) records the WRONG branch → next

  be post rewrites that branch's history; get.js:423-429.

  • JGET-004] CRIT dirty cross-branch switch writes conflict markers into the

  uncommitted file + grows wtlog (case 09); dirtyOverlapCheck get.js:328.

  • JGET-005] HIGH wrong branch fetched for any non-main/master name (case 24);

  pickWant wire.js:107 builds refs/<b> not refs/heads/<b>.

  • JGET-006] HIGH a pre-existing empty .be/ dir crashes jab get (the

  hermetic-shield shape); discover.js find() + get.js:128-134 + seedRemote.

  • JGET-007] HIGH sub fetch by EXACT PIN fails when pin≠advertised tip;

  submount.js:128-138 (no branch-fetch+history-resolve fallback).

  • JGET-008] HIGH sameSourceUri emits malformed ssh:////host;

  submount.js:121 uses u.authority (already //-prefixed) not u.host.

  • JGET-009] HIGH grandchild sub title from the WRONG .gitmodules → HOMENOPROJ

  (cases 31/45); depth-2 reads the parent top-level .gitmodules.

  • JGET-010] MED path-vs-branch misroute (cases 34/33/04); restorePath

  get.js:390 runs before the branch/detach branches (see DIS-017).

  • JGET-011] MED parseRemote srcRoot misderive (cases 29/55); get.js:96-101.
  • JGET-012] MED relative sub url ../sub not resolved vs parent source

  (case 50, SUBS-020).

  • JGET-013] MED --force leaves a tracked orphan (case 49, GET-016).
  • JGET-014] LOW same-line weave: jab cnf/rc=1 vs native mrg/rc=0, and

  ours/theirs SWAPPED (case 66).

TODOs

• Land JGET-002/003/004 (CRITICAL) first, each repro-first + a be/test/get gate (spec-based since TEST-003 — native is no longer the oracle).
• Then HIGH: 005 open (wire.js:122 still refs/<b>), 006 open (seedLocal face), 007 PARTIAL (local half landed d80bdb5c); 008/009 DONE (SUBS-047).
• Then MED (010-013; 011 PARTIAL — the .be-FILE half landed via GET-038/DIS-058), then LOW (014 — premise needs a spec ruling, see the ticket).
• Reconcile DIS-058 (recursion is NOT read-only in the code under test).
• Re-triage 004/014 against GET.mkd ("maybe re-applying uncommitted changes") + be/test/get/dirty (d80bdb5c pins weave + loud-conflict) — the native-refusal/native-mrg oracles are retired.
• Cleanup: native be leaked a .be/srcgit shard into ~/.be during a git-source oracle clone — prune it.

Blockers and bummers

• JGET-006 (empty-.be crash) masks many submodule/clone cases (the native harness pre-creates wt/.be), so it gates clean repros for several children.
• d80bdb5c (TEST-003) retired native be from be/test ("LAGS jab") — every "parity-port vs native" TODO in the children is now a spec-gate TODO instead.

Outcome

• (open) Filed 2026-06-29 from the JS-096 differential investigation.
• 2026-07-07 audit: JGET-008/JGET-009 RETIRED (URI-013 c43324f4 + SUBS-047, unit test be/test/submount.js); JGET-007 local-source half landed (SUBS-046, d80bdb5c); JGET-011 .be-FILE-redirect half landed (GET-038, in the DIS-058 changeset da7fa5a7); the rest re-verified still present in be/verbs/get/get.js / be/shared/wire.js.