DOG-005: graf cut-over to `dog/WEAVE`; delete `graf/WEAVE`

Now: DONE

Repoint graf at dog/WEAVE and delete graf/WEAVE.{c,h}. GET, REBASE, BLAME, DIFFREF and LOG call the new API; the parent-closure WEAVEApply path becomes WEAVEMerge + scope bitsets resolved at the graf/DAG boundary. Finishes the move, updates both INDEX.md, and retires the superseded tickets. Gates on DOG-003/DOG-004.

Input

Context

Callers: graf/GET.c (build_tip_weave*), graf/REBASE.c, graf/BLAME.c (GRAFFileWeave), graf/DIFFREF.c, graf/LOG.c.
GRAFFileWeave folds the closure via WEAVEApply/Diff; this becomes WEAVENext along first-parent + WEAVEMerge at merges, scope from the DAG.
THE bug: be-patch-15 step 2 fails 5/5 under the dog cut-over. DIS-045 PROVED the dog engine correct — dog/test/WEAVE03 drives step-2 as a shared-base 3-way merge (base F2 → ours, base F2 → theirs, then WEAVEMerge) and gets 08.lib.want_step2.c byte-exact (322/322, 4.7M fuzz). So the fault is graf's GLUE: WEAVEApply-against-parent-closure replay is NOT byte-identical to that first-parent WEAVENext+WEAVEMerge shared-base model for the ancestor-skip topology. Fix the modeling, not the engine.

Goals

All graf weave ops on dog/WEAVE; graf/WEAVE.{c,h} deleted; suite green; graf blame/diff/merge + be get/be patch byte-parity (modulo capped).
GET-001's per-version BASS blowup gone by construction (no per-call decode).

Constraints

Behaviour-preserving for uncapped files; conflict-marker output shape unchanged so POST-006/DIS-040 detection is unaffected.

WIP

Design decisions

Resolve closures to scope bitsets at the graf/DAG edge: map a rev's ancestor u64 ids → commits[] indices once per op (reuse DAGAncestors).
Model the ancestor-skip merge as the DIS-045 WEAVE03 shared-base 3-way: fold ours AND theirs onto the SAME merge-base weave (the skip ancestor, NOT a per-branch closure replay), then WEAVEMerge; that is what the engine is byte-parity-correct for. Make GRAFFileWeave's glue match it.
Land file-disjoint where possible; cut over one caller at a time if needed.

TODOs

Repro be-patch-15 step 2 5/5 fail through the graf glue (the dog engine is already pinned by WEAVE03 — isolate the closure→fold model).
Repoint GET/REBASE/BLAME/DIFFREF/LOG; closure→scope at the DAG boundary, shared-base 3-way at ancestor-skip merges (per WEAVE03).
Delete graf/WEAVE.{c,h}; update graf/INDEX.md + dog/INDEX.md.
Full ctest + blame/diff/merge ident; close GET-001/DIS-003/DIFF-007 and re-confirm MEM-036/MEM-040 are moot.

Blockers and bummers

Gate CLEARED (audit 2026-07-07): DOG-003 landed 489865d8, DOG-004 landed cba618d4 (libdog). Cut-over not started: graf/WEAVE.{c,h} still present (5 WEAVEApply sites), graf/GET.c:19 still includes the graf-local WEAVE.h.
Gates on DOG-003+DOG-004. DIS-045 no longer blocks — engine proven correct, the fault is the graf cut-over glue (closure replay ≠ shared-base 3-way). Watch DIS-041 (FF-weave needs indexed commits) during GET cut-over.

Outcome

(open)

DOG-005: graf cut-over to dog/WEAVE; delete graf/WEAVE