(motir-meta) Decide whether the planning corpus is COMPRESSED — partitioning is now measured to cap the hardest pass at a 29% reduction

Repo: motir-meta. One PR. A decision record; it changes no rule text in the same pass.

This card exists because MOTIR-2626 reached a verdict its own scope could not act on. prompts/plan-rules/CLASSIFICATION.md measured the partition against origin/main, and the answer is that partitioning alone takes the hardest planning pass from 2 938 lines to 2 089 — a 29% reduction. The story's premise was that a pass reads what it applies; the measurement says the hardest pass still reads 71% of the corpus, because the great mass of it is universal body-writing prose that no phase, kind or type selector can cut.

Partitioning was still worth doing and is not in question here. The question this card answers is what the SECOND lever is, now that the first one's ceiling is known rather than assumed.

The questions it must settle

1. Is the incident narrative compressed, and by what rule? Most rules carry a fixture — a card key, a date, a quoted failure — and those fixtures are a large share of the bulk. They are also what makes a rule believable and what stops it being re-litigated. Decide the split: what stays inline, what moves to a per-rule fixture file the reader can follow, and what is dropped.
2. Does the per-card gate checklist become an INDEX plus per-pack expansions? MOTIR-2626 kept it whole in core.md (307 lines on every cell) because the alternative needed a re-wording its sibling was forbidden from making. That constraint does not bind this card. The checklist is cited BY NUMBER from across the corpus and from run.md, so any re-shape must preserve gate numbering.
3. Is core.md capped? It is 688 lines (23.4%) and arrives on every pass, which makes it the one pack where a line costs the most. Decide whether a budget applies and what happens when a rule wants in.
4. What is the target, in lines, for the worst cell — so the next measurement has a bar to be judged against rather than a direction.

Acceptance criteria

1. A decision record exists at prompts/plan-rules/COMPRESSION.md, structured Status → Context → Decision → Consequences.
2. Each of the four questions is answered explicitly; anything left open names the card that will settle it.
3. The decision cites CLASSIFICATION.md's measured numbers, re-derived on the corpus as it stands at that time — not the figures quoted above, which are a reading of commit f7d5723.
4. Whatever rule it fixes for the fixtures is stated so that a later pass can apply it to a rule WITHOUT re-deciding per rule.
5. The RULES-vs-LESSONS contract is addressed head-on: a rule fires every time its situation occurs, so a compression that makes a rule harder to recognise is a deletion wearing a diff.
6. Gate numbering survives whatever question 2 decides, and the decision says how.

Context refs

• prompts/plan-rules/CLASSIFICATION.md — the measurement, and its "the numbers the decision is made on" section.
• prompts/plan-rules/DECISION.md — the partition contract this one extends; its question-6 disposition of the checklist is what question 2 reopens.
• prompts/plan-rules.md — the corpus.