Motir

Repo: motir-meta. One PR. The first execution card under MOTIR-2642's COMPRESSION.md. It builds the mechanics the other three compression cards lift, and applies them to the single most expensive region of the corpus.

COMPRESSION.md decides the rule; it deliberately cuts nothing. This card cuts, and the region it cuts is the one where a line costs the most: core.md's per-card gate checklist — 307 lines, 43.9% of core.md, delivered to every planning pass of every phase, kind and type.

Three deliverables, and the first two are what the siblings lift

1. prompts/plan-rules/fixtures/ — one file per pack, listed in MANIFEST.md with the disposition CORPUS-MAINTENANCE.md already has: loaded by NO planning pass. A fixtures file a pass loads saves nothing; its reader is the rule-authoring path and anyone re-litigating a rule.
2. The evidence tag, in the fixed form COMPRESSION.md § Decision 1 states, applied to every one of the 16 numbered gates that carries warrant.
3. A guard on split.py — see below. It is in this card because this card is the one that makes the hazard live.

⚠️ split.py becomes destructive the moment this card lands

split.py regenerates every pack from the pre-split corpus at f7d5723. Today that is a safe no-op — re-running it reproduces the packs byte-identically. After this card's compression it does not fail, it silently reverts the corpus and prints success. verify.py at least goes red; this one does not. Fixing it is part of this card because this card creates the exposure.

The rule being applied (do not re-derive it — read § Decision 1)

A rule is TRIGGER + DIRECTIVE + WARRANT. Only WARRANT moves. The discriminator is mechanical: delete the span and ask whether the rule still FIRES. Changes when it fires ⇒ trigger, it stays, whatever incident it cites. Changes only whether a reader BELIEVES it ⇒ warrant, it moves.

Gate 1's lexical tells ("a second repo's FILE or SYMBOL named anywhere inside an acceptance criterion") are TRIGGER and stay inline in full. Gate 1's "Three instances in two days … (MOTIR-2057, notes.html #215 … MOTIR-1983, #216)" is WARRANT and moves.

Acceptance criteria

1. prompts/plan-rules/fixtures/core.md exists and holds, verbatim, every warrant span this card removes from prompts/plan-rules/core.md, each under an anchor the inline tag points at.
2. MANIFEST.md lists fixtures/ in its "The files" table with loaded when = never — no planning pass loads these, in the same words CORPUS-MAINTENANCE.md carries. The row is added in split.py, which generates that table — a hand-edit to MANIFEST.md alone drifts on the next regeneration.
3. Every one of the 16 numbered gates that carried warrant carries exactly one evidence tag in COMPRESSION.md § Decision 1's four-slot form, and no gate's trigger or directive text is re-worded.
4. Gate numbering is untouched — 16 gates, same numbers, same order, one runnable list in core.md. No gate moves to another pack (§ Decision 2 is settled; this card does not reopen it).
5. python3 prompts/plan-rules/COMPRESSION.measure.py reports the checklist region below 200 lines, down from 307. AMENDED 2026-08-11 during this card's run — the number was unreachable under § Decision 1, and the target gives way, not the rule (§ Decision 4). The replacement: every attributable warrant span in the region is moved and the residue is REPORTED with its arithmetic — COMPRESSION.measure.py prints the region's new size, and the PR body states what stayed and why.
   • The evidence. COMPRESSION.md measures the corpus-wide warrant share and takes the SENTENCE grain as its working figure (29.9%). This criterion carried that proportion across to a region it had never been measured on. The record's own classifier, scoped to core.md lines 31–337, returns 103 lines of 307 — so 307 − 103 = 204, above 200 before one tag line is inserted, and § Decision 1 mandates one tag per compressed rule. The only grain that reaches under 200 is the BLOCK grain (258 lines), which removes triggers and directives along with the evidence.
   • Delivered: 307 → 278, 4 258 characters of warrant moved to fixtures/core.md in 19 spans across 5 gates. The gap to 103 is that the classifier marks any sentence containing a citation, which sweeps in pure triggers that happen to cite.
   • Amended per run.md's calendar rule: COMPRESSION.md merged at 14:29, this card was authored at 14:00–14:07, so the record outranks the card and a contradicted clause is amended on the record rather than dropped. Planning bug: MOTIR-2687. Lesson: notes.html #258 — renumbered from #257 when motir-meta PR #170 merged first and took that number; #257 (MOTIR-2686) is the sibling instance of the same family, and #258 cross-links it.
6. WARRANT CONSERVATION is proved in the PR body (COMPRESSION.md § Decision 1, invariant C): the diff against the pre-compression tree contains only (a) deletions whose text appears verbatim in fixtures/core.md and (b) inserted tag lines matching the fixed form. Nothing else.
7. split.py refuses to overwrite a compressed pack: it compares what it would write against what is on disk and exits non-zero naming the divergent packs, unless --force is passed. Proved by running it and showing it declines.
8. prompts/plan-rules.md, DECISION.md and CLASSIFICATION.md are not edited — they are records of what was measured and decided, not of the corpus's current line counts.

Context refs

• prompts/plan-rules/COMPRESSION.md — § Decision 1 (the tag form + the deletion test), § Decision 2 (why the checklist is not split), invariant C, and the split.py hazard in § Consequences.
• prompts/plan-rules/COMPRESSION.measure.py — the measurement, re-runnable; --check asserts the caps.
• prompts/plan-rules/split.py — the generator; it authors MANIFEST.md and would revert the packs.
• prompts/plan-rules/core.md lines 31–337 — the checklist.