DIS-044: WEAVEMerge RGA order key is not causally monotonic

DIS-043 gave dog/WEAVE tokens an RGA stable order and sorts concurrent siblings by raw commit-id DESC. That is only causally valid when a commit's id is larger than its ancestors' — true for the DWEAVE fuzzer (it keys commit-id = line index, monotonic) but FALSE for real 60-bit commit hashlets, which are arbitrary. When the common-ancestor (spine) id exceeds the edit ids, a replace-edit's inserted token is stranded behind the dead original's whole subtree, and a same-line divergence can splice clean instead of conflicting — silent merge data-loss. Found during DOG-005 (graf cut-over): diff/blame/cat/get are byte-parity-clean, but 11 real-commit merge tests fail. The green DIS-043 fuzz was misleading — its monotonic-id oracle could never hit this.

Issues

RGA needs a causally-monotonic order key; the raw commit-id is not one.

Repro (isolated dog-API harness): the bug toggles purely on base<ours<theirs
(passes) vs base>ours (fails) — int x = 10 merges to int x = ; … 10 at EOF.
Failing: be-patch-03/04/10/12/14/15/17/22/23/34, be-get-51-ff-weave-unindexed-merge
(real-commit merges with arbitrary hashlets); synthetic small ordered ids (1/2/3) hide it.
Root: a descendant commit may have a SMALLER hashlet than its ancestor, so commit-id
order ≠ causal order; RGA's concurrent tie-break must be a causal/Lamport rank.
Fuzzer gap: dog/fuzz/WEAVE keys commit-id = line index (monotonic), accidentally a
valid causal order — so it can never surface this class.

Planned

Tie-break on the commit INDEX (already stored, causal), made path-independent.

The token already stores the commit INDEX, and commits[] is built ancestors-first
(spine = index 0), so index(ancestor) < index(descendant) — causal. Switch WEAVEMerge's RGA tie-break from the raw hashlet (cid, dog/WEAVE.c:519/588) to the index; that fixes the spine-id > edit-id stranding directly. No new column.
CAVEAT — make the index path-independent: WEAVEMerge builds commits[] as
a.commits ++ b.new, so concurrent siblings flip index by parent role (the very criss-cross DIS-043 used the hashlet to avoid). Build the merged commits[] in a deterministic causal order (topo: ancestors first, concurrent commits tie-broken by hashlet) so the index is a real topo rank — causal AND path-independent.
Extend dog/fuzz/WEAVE to non-monotonic / arbitrary commit-ids (its commit-id =
line-index oracle is why DIS-043 looked green); fold the DOG-005 base>ours repro + the failing be-patch/be-get merges in as property tests, confirmed red pre-fix.

Blockers

DIS-043 is LANDED (8eece3f3) and carries this bug; this fixes it. DOG-005's
merge cut-over gates on it (the diff/get cut-over is already byte-parity-clean).

Landed

(open)