Beagle SCM

JGET-004 CRIT: dirty cross-branch jab get ?feat writes conflict markers into the uncommitted file (no up-front dirty refusal)

jab get ?feat against a worktree with a DIRTY tracked file (locally edited, uncommitted) does not refuse: its only dirty guard fires solely for NEW-in-target paths that lack a baseline, so a tracked file present in both trees falls straight into the leaf 3-way weave. The on-disk dirty bytes vs base vs target collide, jab WRITES <<<<dirty||||-feat>>>> into the file and grows the wtlog, then errors GETCONF. The user's uncommitted bytes are mangled and the refusal-atomicity invariant is violated. Native refuses up front (SNIFFDRTY), leaving the file and wtlog untouched. Part of JGET-001; method Issues.

Input

Context

• Native beagle/test/get/09-refuse-no-commit (cross-branch-dirty refusal,

  SNIFFDRTY, must run BEFORE the commit point).

• Repro: two-branch repo, edit tracked x.txt dirty, jab get ?feat → x.txt

  becomes <<<<dirty||||-feat>>>>, wtlog 7→8 rows, then rc=1 (GETCONF). VERIFIED marker in evidence /tmp/claude-1000/jabprobe/r09/wt/x.txt. Native: rc≠0, x.txt still dirty, wtlog unchanged.

• Root: dirtyOverlapCheck (get.js:328) only inspects paths with NO baseline

  (new-in-target); a tracked file in both base+target skips it and reaches the leaf weave-merge (~get.js:651/748).

Goals

• A switch/restore that would overwrite a DIRTY tracked file refuses up front

  (before any wt write or wtlog append), wt + wtlog byte-identical, nonzero exit with a native-style dirty diagnostic.

• A clean wt still switches/weaves as today.

Constraints

• The refusal must be a PRE-PASS barrier (before the fan-out commit point), like

  native SNIFFDRTY / the existing SNIFFOVRL pre-pass. Repro-first.

WIP

Design decisions

• Broaden the pre-pass: refuse if ANY tracked path that the target changes is

  dirty on disk (baseline≠wt), not only un-baselined overlays. Keep the existing no-baseline overlay arm. Decide whether --force/weave is the opt-in to merge.

TODOs

• SPEC RULING FIRST (gritzko): native SNIFFDRTY refuses; GET.mkd says GET resets "maybe re-applying uncommitted changes", and be/test/get/dirty (d80bdb5c, TEST-003 — native retired as oracle) PINS the weave + loud-conflict (rc≠0 + markers) behavior — a refusal fix would contradict that suite. Decide refuse-vs-weave for the cross-branch-dirty case.
• If refuse wins: widen dirtyOverlapCheck (get.js:607-632 — still baseline-less paths only; a tracked dirty file still reaches the leaf weave get.js:767-786) into a pre-pass over baselined-but-dirty target paths; repro test = dirty tracked file + ?feat → no write, no wtlog growth, nonzero.
• Either way fix the ATOMICITY half: the wtlog tip row is appended BEFORE the fan-out (get.js:516), so an erroring get still grows the wtlog — same defect SUBS-047 flagged as an own-ticket candidate.

Blockers and bummers

• Destroys uncommitted user edits; pair-review with JGET-002 (both are

  "missing pre-flight refusal" in the seed).

• AUDIT 2026-07-07: premise partly superseded — the marker-writing weave is now arguably the SPEC'D behavior (GET.mkd + test/get/dirty case A2); only the wtlog-growth/atomicity part is an undisputed bug. Left open pending the ruling.

Outcome

• (open) Filed 2026-06-29.