Beagle SCM

GET-057: bare get CLOBBERS dirty staged files untouched by the incoming commit

Incident 2026-07-29 21:31, gritzko in //PATCH-025 (the wt runs its OWN modified get.js — PATCH-025 rewired the weave leaf, so trunk may be innocent): be get absorbed trunk's ULOG-004 commit (da3cf823 → 543269d6, touches only shared/wtlog.js + test/wtlog_con.js + CMakeLists) and OVERWROTE five dirty+staged test files with trunk bytes although theirs==base for every one of them — keep-ours was the only legal outcome. Staged demotion ...V → ...v also observed (put rows precede the new get tip row; rows NOT deleted). Regold content lost from the tree (put rows carry no blob sha; staging is row-only).

Input

Context

• Clobbered (byte-identical to trunk after the get): test/get/conf-state/run.sh, test/get/confdrop/run.sh, test/get/dirty/run.sh, test/post/conf-state/run.sh, test/status/conflict/run.sh.
• Survived (kept local bytes, restamped 21:31): test/patch/{conflict,dirty-overlap,same-anchor-conflict}/{run.sh,golden.out}, test/patch/stacked-conflict/* (untracked-new), test/get/detached-recover/run.sh, test/wtlog_con.js. All were equally staged+dirty — the discriminator is UNKNOWN, that is the first thing to find.
• Ulog tail of //PATCH-025/test after the incident: get /test seed, 14 put rows (20:17 and 21:20 batches), then the new get /test row — no rows lost.
• The wt's jab loaded the wt's OWN jsrc: verbs/get/get.js there carries PATCH-025's changes (weave3 → mergedLive {bytes,spans}, POST-032 unresolved-conflict re-weave skip REMOVED, carrySweep con band row-based). Evidence tree frozen: do NOT edit //PATCH-025 until the repro is extracted.

Goals

1. Reproduce in a scratch clone: stage+edit files upstream-untouched by the incoming commit, bare get, diff. Run once with the wt's get.js and once with trunk's — pin WHICH code clobbers, and what separates the clobbered five from the survivors.
2. Fix so theirs==base ⇒ keep ours, always — staged or not; add the repro as a permanent get test.
3. RULED (gritzko 2026-07-29): an FF get keeps the staged list — recorded in GET-053 with the FF-only redesign; this ticket owns only the clobber.

Constraints

• If the bug is PATCH-025's, the fix belongs in that ticket's scope before it can land; if trunk's, this is a standalone HIGH.
• Recovery of the five regolds is mechanical (itemised in PATCH-025 Outcome + worker report) — redo only AFTER the repro is captured from the frozen tree.

WIP

Design decisions

• DISCRIMINATOR = THE EXEC BIT. The clobbered five are exactly the mode 100755 files; every survivor is 100644. Verified by stat on the base tree.
• Not the ulog batches, not the stamp bands, not carrySweep, not weave3: the 20:17/21:20 put batches straddle both sets (get/detached-recover, 21:20, survived).
• ROOT CAUSE: shared/checkout.js apply() GET-040 dirty guard was gated on leaf.kind === "f", so a dirty EXEC ("x") leaf fell through to clean-overwrite.
• Only submount.mount calls checkout.apply (submount.js:365,421), so the clobber is SUBMODULE-ONLY — a flat wt weaves through get.js's own leaf, which is correct.
• That is why the incident hit //PATCH-025/test (a sub) and spared the three staged parent .js files: 644, and carried by carrySweep (band mrg, ceil-1ms).
• GET-056 fixed this SAME kind === "f" gate in get.js's weave leaf (test/get/dirty-exec-binary) but missed the checkout.js twin — this is that miss.
• TRUNK IS GUILTY. PATCH-025 is INNOCENT: the replay clobbers the identical five under da3cf823's own code and under the evidence code — byte-identical outcome.
• weave3's theirs==base ⇒ keep ours arm is present and correct in BOTH; it is never reached for a sub file, which is the whole bug.
• mtime forensics confirm the branch: survivors 21:31:56.248 (mrg band = ceil-1ms, restamped only), clobbered .249 (ceiling = the clean-overwrite stamp).

TODOs

• discriminator: why those five, not the patch/* files — the exec bit
• two-code repro (wt get.js vs trunk get.js) — both clobber, trunk guilty
• keep-ours fix + permanent test (test/get/sub-dirty-exec)
• redo the 5 regolds in //PATCH-025 after evidence capture → done, see PATCH-025 Outcome
• gritzko's ruling on staging surviving a base advance → GET-053 (keep the staged list)

Outcome

• CLOSED via GET-056b landing ///be/#5c0d5c21 + //be/test#4e59c1ac (2026-07-30): the same gate fixed to the full merge matrix, equivalent test test/sub/dirty-exec-keep landed.
• //GET-057's staged one-liner + test/get/sub-dirty-exec/run.sh SUPERSEDED, never landed; the worktree holds nothing unique (evidence copies in //GET-057/evidence/) — reapable.
• THE FIX (trunk-only, standalone HIGH; no PATCH-025 lineage needed — PATCH-025 never touched checkout.js): checkout.js:203 leaf.kind === "f" → (leaf.kind === "f" || leaf.kind === "x"). Symlinks keep clean-resetting (a dirty symlink is a re-point, out of scope).
• TEST be-js-get-sub-dirty-exec: parent+sub, sub gains an exec run.sh + a 644 note.txt, both dirtied AND put-staged in the mounted sub, sub advances on an UNRELATED file, update get. Asserts ours kept for both, the exec bit kept, and that theirs still fast-forwards. RED on the old gate ("dirty+staged EXEC run.sh was CLOBBERED"), GREEN after.
• INCIDENT REPLAY (the real evidence, not a toy): clone ///be/#da3cf823, stage+dirty the same 14 test files, bare jab get. Trunk code → the exact five clobbered, nine kept. Evidence (PATCH-025) code → identical. Fixed code → lost=0, and theirs' test/wtlog_con.js still lands.
• ctest -j16 = 318/318 PASS (get+patch+status+post = 132/132). Pre-existing suite was 317; the new case is the 318th.
• Evidence copies of the frozen wt's three JS files kept in //GET-057/evidence/; repro scripts in //GET-057/scratch/ (untracked, not staged).
• NOT DONE (out of scope, by instruction): GET-053's FF gate. The ...V → ...v staged demotion reproduces in a FLAT wt too and is GET-053's ruling, not a clobber.