Beagle SCM

DIS-045: dog/WEAVE merge mis-anchors a blank line vs a code-line EOL

be patch step 2 of test/patch/15-ancestor-skip merges lib.c to the WRONG content on dog/WEAVE — the DIS-003 "insert-vs-blank" case: a re-stamped spine blank line and theirs's code-line EOL share their bytes (\n), so the content-hashlet diff/merge anchors the inserted line on the wrong side of the blank. It was a ~½ FLAKE on the old graf engine (DIS-003); under the DOG-005 cut-over it is a HARD 5/5 failure, so dog is not byte-parity with graf for this case. Distinct from DIS-043 (criss-cross) and DIS-044 (causal order) — this is an identical-CONTENT token ambiguity, not an ordering one. It is the last case blocking DOG-005's merge cut-over (diff/blame/cat/get are already byte-parity-clean).

Issues

Identical-content tokens (blank vs EOL) are not disambiguated at a merge boundary.

• Repro: test/patch/15-ancestor-skip step 2 (08.lib.want_step2.c) — merged lib.c

  differs (deterministic, 5/5). Inputs are in that dir (base/ours/theirs .c).

• Root (DIS-003 hypothesis, confirm): a blank line \n and a code-line's trailing

  \n hash equal, so WEAVENext's BRAM/NEIL LCS (or the WEAVEMerge join) pairs the wrong one and the insert lands on the wrong side of the blank.

• The trunk's be-patch-15 is still GRAF-backed (DOG-005 unlanded), so reproduce via

  the dog API directly (feed the step-2 base/ours/theirs through WEAVENext/WEAVEMerge/WEAVEProduce), not the graf-backed test.

Planned

Disambiguate identical-content tokens; make the fuzzer inject them.

• Find where the blank/EOL pair mis-anchors (WEAVENext diff vs WEAVEMerge join) and

  fix so the merge produces the correct content; preserve DIS-043/DIS-044.

• Extend dog/fuzz/WEAVE to inject blank lines / repeated identical-content tokens at

  merge boundaries so this class is caught; fold the step-2 case in as a property test (confirmed red pre-fix).

Blockers

• No longer blocks DOG-005. The premise (engine mis-anchors blank vs EOL)

  is FALSE: DIS-043 (idh JOIN) + DIS-044 (causal order) already removed both DIS-003 layers, so the blank-vs-EOL precondition cannot arise in the post-DIS-044 engine. The be-patch-15 5/5 fail is graf cut-over GLUE, not the engine — re-tasked into DOG-005.

Outcome

• Close: ALREADY FIXED by DIS-043/DIS-044; engine pinned, fuzz widened,

  NO repro. No engine change — the changeset is pure teeth: dog/test/WEAVE03 (patch/15 step-2 shared-base oracle, 08.lib.want_step2.c byte-exact), WEAVE01 dis045_insert_blank/dis045_double_blank criss-cross-over-blank cases, and dog/fuzz/WEAVE W2_BLANK injection. 322/322 green; DIS-043/044 not regressed; ~4.7M fuzz execs with blank injection found no mis-anchor.

• Committed 711618fe ("DIS-045 WEAVE: blank/EOL merge teeth + step-2

  oracle"); the be-patch-15 repro lives on in DOG-005.