Motir

Repo: motir-meta. One PR. The last execution card under MOTIR-2642's COMPRESSION.md, and the one that closes the target.

AMENDED 2026-08-11 by this card's own run, before a line was written — AC 3's target was 1 600 and is now 1 950. MOTIR-2706 merged (motir-meta PR #174) at 22:39Z on 2026-08-11, eight hours AFTER this card was authored (14:01Z), and re-set § Decision 4 from a projection to a measurement. Under run.md's a done blocker that merged AFTER the card outranks the card, the record wins and the criterion is amended rather than quietly dropped. This is also MOTIR-2702's AC 3 discharged — "MOTIR-2662's own AC is reconciled against the corrected number BEFORE it is run."

The rung-2 evidence, read off the amended record and re-measured on origin/main at 0d2a64c: the worst cell is 2 013 today, not 2 188 (MOTIR-2660/2661 landed in between). A perfect, lossless, two-line-tag cut of the six packs this card still owns floors it at 1 867 — so 1 600 was 267 lines below anything achievable and was never available. The three landed compression cards realised a mean 38% of their projections, which puts the expected landing at 1 948. § Decision 4 sets 1 950 there deliberately: "where the work is expected to land, not where a perfect cut would."

The twelve packs no sibling card takes: kind-story (215) · phase-skeleton (182) · op-replan (167) · kind-leaf (122) · type-design (118) · type-migration (111) · kind-bug (91) · type-code (58) · type-test (55) · type-bars (54) · kind-container (50) · op-sprint (47). 1 270 lines, 358 of them warrant. (Re-measured 2026-08-11 on 0d2a64c: both figures still hold exactly.)

Only six of the twelve are in the worst cell — op-replan · kind-leaf · type-migration · kind-bug · type-code · type-bars, 603 lines. The other six are real corpus savings that AC 3 cannot see, which is why AC 4 exists.

op-replan.md is the densest pack in the corpus by warrant share — 49.1%, nearly half — and it is an overlay, so it lands on top of an already-expensive cell. It is the largest single win here.

type-code.md is the leanest in the corpus at 8.6% — five lines of warrant in fifty-eight — and op-sprint.md is at 4.3%. Both are expected to come out close to unchanged. That is a result, not an omission: state it in the PR body rather than manufacturing an edit.

Acceptance criteria

1. Every warrant span removed appears verbatim in the matching prompts/plan-rules/fixtures/<pack>.md, under the anchor its inline tag points at.
2. Each compressed rule carries exactly one evidence tag in COMPRESSION.md § Decision 1's four-slot form, and no rule's TRIGGER or DIRECTIVE is re-worded.
3. The worst legal cell — deepen · bug · code under a re-plan — is at or under 1 600 1 950 lines (§ Decision 4's target as amended by MOTIR-2706), down from 2 188 2 013, as reported by python3 prompts/plan-rules/COMPRESSION.measure.py --check. § Decision 4's escape clause is re-affirmed and NOT spent: if the cell lands above 1 950 without cutting a trigger, a tell-list or a shape table, the target gives way again and the miss is reported with its arithmetic — the form MOTIR-2661 already used on its own AC.
4. Any pack left unchanged is named in the PR body with its measured warrant share, so a reader can tell "nothing to compress" from "not reached". All twelve are accounted for either way.
5. WARRANT CONSERVATION is proved in the PR body (invariant C): the diff contains only deletions whose text appears verbatim in the matching fixtures file, plus inserted tag lines matching the fixed form. Run it as COMPRESSION.conserve.py --base <pre-compression ref> — the default merge-base(HEAD, origin/main) baseline is only correct from inside an unmerged PR (MOTIR-2711).
6. DECISION.md, CLASSIFICATION.md and plan-rules.md are not edited.

Context refs

• prompts/plan-rules/COMPRESSION.md — § Decision 1 (the tag form + the deletion test), § Decision 4 as amended (the 1 950-line target and how it is asserted, and the clause that says the target gives way before a trigger does), invariant C.
• prompts/plan-rules/COMPRESSION.measure.py --check — the assertion; CORE_CAP = 620 PASSES at 598, WORST_TARGET = 1950 fails until this card lands, by design.
• prompts/plan-rules/fixtures/ — the tree MOTIR-2659 created.
• ⚠️ COMPRESSION.measure.py's proj column SIZES the problem and does not set a floor — § Consequences' three defects. Its TAG_LINES = 1 still contradicts § Decision 1's two-line tag (MOTIR-2702, open), so every projected figure is optimistic by one line per warranted rule.