Motir

Repo: motir-meta. One PR. § Decision 3 says it in its own words: "Raising the cap is a decision record, not a commit. It amends this one, with a measurement." MOTIR-2660 produced the measurement. This card is the amendment.

What was measured

MOTIR-2660 compressed core.md to 598 lines (670 in, 699 before any compression) with invariant C green — and stopped, because the remaining distance to § Decision 3's 520 cap can only be closed by cutting triggers, tell-lists or the → index.

The cap was derived from COMPRESSION.measure.py's projection of 501. That projection is total − sentence-grain warrant + one tag per warranted rule, and its warrant detector is a regex for citation markers (MOTIR-\d+, a date, notes.html #n, #NN, fixture). Two defects follow, both counted rather than argued:

1. It scores TRIGGER and DIRECTIVE prose as warrant whenever that prose cites a key. Of the 22 marker-bearing sentences left in core.md outside the evidence tags, none is warrant under § Decision 1's deletion test — 9 are rule HEADINGS whose attribution is part of the rule's name, 6 are TELLS, 7 are DIRECTIVES carrying a → pointer or a scope clause. § Decision 1 names this hazard itself: "Gate 1's … list of things to look for … Gate 14's ORDERING axis … Both cite incidents; neither is warrant."
2. It is self-referential. An inserted evidence tag starts with >, carries a key and a date, and scores as compressible warrant — so the projection drifts UPWARD as warrant is removed: 501 before MOTIR-2659, 542 after it, 546 mid-way through MOTIR-2660.

What this card decides

Amend COMPRESSION.md — a Status/Context/Decision/Consequences record amendment in the style the file already uses, struck-through rather than rewritten where a number changes, so the original decision stays readable.

Acceptance criteria

1. § Decision 3's core.md cap is re-set from the measured floor, with MOTIR-2660's 598 and the 22-sentence enumeration quoted as the measurement; the "a new rule PAYS FOR ITSELF" mechanism is kept — it is the half of § Decision 3 the measurement does not touch.
2. § Decision 4's 1 600-line worst-cell target is re-stated against the same evidence, or explicitly left standing with the reason; MOTIR-2662 is judged against whatever this card leaves.
3. The projection's two defects are recorded in § Consequences — the marker regex scoring triggers, and the self-reference — so the next reader does not re-derive the caps from the same number.
4. COMPRESSION.measure.py's CORE_CAP / WORST_TARGET constants match the amended record, and --check reports what the record says.
5. DECISION.md, CLASSIFICATION.md, core.md and the fixtures are not edited — this card changes the RECORD, never a rule.
6. The amendment carries its own date and cites MOTIR-2660 as the run that produced the measurement.