Ratchets compare a STALE merge ref against origin/main's LIVE tip ('no merge-base available; stricter') — a main-side baseline shrink reads as GREW on every in-flight PR (cb200 601→599 tonight; PMAT-3539 'deleted' earlier) #3643

What

Every differential ratchet in guard-tree (check_baseline_ratchets.sh, check_complexity_ratchet.sh, check_roadmap_diff_additive.sh, …) prints on a PR run:

comparand: tip of origin/main (no merge-base available; stricter) (protected; a pull request cannot rewrite it)

"Stricter" is wrong in one direction: the PR's checkout is the merge ref built when the run was triggered, and the comparand is origin/main fetched at guard time. When main moves between the two — tonight #3491 shrank scripts/cb200_baseline.txt 601→599 and the complexity baseline — every PR whose run was in flight reads as a violation of what it never touched:

• PMAT-3459 + PMAT-3460: register the in-repo tag-path ticket (P0, 0.69.0 must-carry) and the epic-by-number follow-up #3461 (roadmap-only + receipt) and #3605: removed_by gets a shape — refusal-receipt-v1, with a closed sentinel set so the required field manufactures nothing #3613: FAIL ratchet scripts/cb200_baseline.txt GREW. It is SHRINK-ONLY vs 237fbc32f2 (01:00Z).
• Earlier the same night, check_roadmap_diff_additive.sh on PMAT-3459 + PMAT-3460: register the in-repo tag-path ticket (P0, 0.69.0 must-carry) and the epic-by-number follow-up #3461 / docs(CLAUDE.md): workspace-test is a 3-shard matrix with a fan-in check; mini-m4 has a lane (PACK-001, #3488) #3494 / feat(release): APR-RELEASE-001 §5 P0·Instrument — make build-report reads the 1092 ledger records nothing read #3271: VIOLATION deleted: 1 base id(s) missing at head: PMAT-3539 — main had gained an entry the merge ref predated.

Cost: one full CI round per affected open PR per main merge (gh pr update-branch), on a fleet that is the bottleneck; and a reader who trusts the word GREW files a defect against a PR that changed nothing.

Fix

The merge ref carries its own base: on refs/pull/N/merge, HEAD^1 is the main commit the merge was built on. Compare against that (git rev-parse HEAD^1, available even at fetch-depth: 1 if the parent is fetched — or fetch-depth: 2), and print it as comparand: merge-ref base <sha>; fall back to the origin/main tip only on a push event (where HEAD IS main) and say so. Then "GREW" means the PR grew it.

done_when

• every ratchet prints comparand: merge-ref base <sha> on a pull_request run and comparand: origin/main tip only on push; the case table has both rows
• mutation: with main advanced past the merge ref by a baseline shrink, a PR that did not touch the baseline is GREEN (RED today); a PR that grew it is RED (unchanged)
• check_roadmap_diff_additive.sh uses the same comparand (its "deleted" of PMAT-3539 on PMAT-3459 + PMAT-3460: register the in-repo tag-path ticket (P0, 0.69.0 must-carry) and the epic-by-number follow-up #3461 was this class)

Not a cut input; it is a CI-round tax on every PR after every merge, which at 20 armed PRs tonight is the difference between the queue draining and not. 0.70.