Motir

Repo: motir-ai. One PR. The mirror half of MOTIR-2969, which fixed the same defect in motir-meta's COMPRESSION.conserve.py (PR #238). Separate card because it is a separate repo — ONE SUBTASK = ONE REPO = ONE PR.

The defect

tests/helpers/corpusConservation.ts is a TS port of classify / isAddition, with one stated deviation: the diff unit is a WORD, not a character. MOTIR-2969's defect has two limbs and the deviation reaches exactly one of them.

Measured against motir-ai origin/main @ c49c6eb, by running the real port:

=== the kind-story limb — a run starting inside a shared `**` marker ===
    violations TODAY: 0      ← UNREACHABLE in the port, as its header claims

=== the phase-deepen limb — a run ending `. **⚠️` ===
  op=insert  isAddition=false
    run head="The prospective limb disciplines what the"
    run tail="the card predicts. **⚠️"
    violations TODAY: 1      ← REPRODUCES UNCHANGED

• Clause (3) is genuinely structural at word granularity. The port drops requirement (3) on the reasoning that an edit landing inside a word can only surface as a replace of that word, which is already fatal. A shared ** prefix is part of a word token (**The vs **For), so the differ has nothing to consume as equal. Confirmed, not assumed.
• Clause (1) ports byte-for-byte, and so does its gap. const SENTENCE_END = /[.!?:][\s"'*)]]*$/;is character-identical to the Python, and the trailing class does not admit⚠️. A run whose last token is **⚠️` fails clause (1) however the words are aligned.

Why it is latent and not red — and why that is the reason to fix it now

classify(BASELINE, SHARED_PLANNING_RULES) on origin/main returns 0 additions, 0 violations. Nothing is broken today. But the constant carries three ⚠️ markers, and the shape that trips this is "a rule inserted immediately before one of them" — which is the single most common edit this corpus receives. The motir-meta side spent three cards (MOTIR-2773, MOTIR-2934, MOTIR-2966) paying down exactly this class one pack at a time before MOTIR-2969 fixed the generator. The mirror is at card zero of that sequence and the fix is already written.

And the mirror has a tempting wrong escape the packs do not have. Its baseline is a refreshable golden copy, so a red clears with pnpm tsx scripts/refresh-planning-rules-baseline.ts. That command is correct for a deliberate re-wording and catastrophic for this one: refreshing past a benign alignment failure also refreshes past any real re-wording sitting in the same diff, and the fixture diff a reviewer is supposed to read would be dominated by the block that triggered it.

The repair — repair (A), and it is already validated at this granularity

MOTIR-2969 chose (A) canonicalise the run before classifying it over (B) widen the clauses, and named the refusal in COMPRESSION.md § Decision 1. Port (A), not (B):

• An insertion's boundary is not a property of the text. Removing b[j1:j2] and removing b[j1-1:j2-1] leave the identical residual whenever b[j1-1] === b[j2-1], symmetrically to the right — over WORD tokens here exactly as over characters there.
• (B) is the limb with the mirror obligation, which is the argument against choosing it in either repo: admitting ⚠️ into SENTENCE_END means the same character lands in both files or the two classifiers diverge on the one clause they still share. (A) leaves SENTENCE_END untouched in both.

Measured, with a word-level alignments() spliced into this port:

phase-deepen limb — run ends `. **⚠️`         violations TODAY: 1   with alignments(): 0
kind-story limb — shared `**` opener          violations TODAY: 0   with alignments(): 0
SAFETY — clause appended inside a sentence    violations TODAY: 1   with alignments(): 1
SAFETY — one word of surviving text altered   violations TODAY: 1   with alignments(): 1

Acceptance criteria

1. classify(BASELINE, SHARED_PLANNING_RULES) is run and quoted BEFORE any edit, on a pristine origin/main worktree: 0 additions, 0 violations — so the fix is measured against a green mirror and cannot be mistaken for clearing a real red.
2. alignments() + asAddition() are ported into tests/helpers/corpusConservation.ts over WORD tokens, mirroring COMPRESSION.conserve.py's functions of the same names, and classify reports the accepted alignment rather than the raw run. SENTENCE_END, SENTENCE_START and MIN_WORDS are unchanged — a diff that touches any of them has chosen refused form (B) and must say so.
3. The port's header block records the asymmetry on the record: clause (3) is structural at word granularity, clause (1) is not, and the alignment window is what closes the second — so the next reader does not re-derive it from the Python.
4. tests/planningRuleConservation.test.ts gains a case for the . **⚠️ shape asserting it is an ADDITION, plus a companion asserting that a word of SURVIVING text altered in the same region is still fatal — the same pairing MOTIR-2969 added to SELF_TEST / ALIGNMENT_GUARD.
5. conserve.py's SELF_TEST table is re-run against the port and every case returns the same verdict, which is the standing evidence that the word-unit deviation is representational and not semantic. The table is now 11 cases, not nine — pull the current one.
6. pnpm vitest run tests/planningRuleConservation.test.ts is green and quoted, and the baseline fixture is NOT refreshed by this PR (git diff --stat shows tests/fixtures/sharedPlanningRules.baseline.txt untouched).

Repeat-defect discharge (plan-rules/kind-bug.md, gate 9)

Six non-archived bug cards now sit under MOTIR-1465 against the conservation checker — MOTIR-2773, MOTIR-2934, MOTIR-2966, MOTIR-2967, MOTIR-2969, and this one — well past the threshold of three, so limb (b) is owed rather than three independent root causes. This card IS the re-shape card's second half. MOTIR-2969 made the (b) argument for motir-meta in its own words — "the corpus does not need a third sweep card; it needs the generator fixed" — and fixed the generator there. The three that preceded it were per-pack symptom sweeps of ONE root shape (an alignment-dependent verdict); MOTIR-2967 is a different root (no baseline existed at all) and does not belong to the count on shape. This card carries the same generator fix across the open-core boundary, which is what makes the family closed rather than paused.

Context refs

• tests/helpers/corpusConservation.ts — isAddition (two clauses), SENTENCE_END, diffWords, classify. classify hands the run straight to isAddition at the same line the Python did.
• tests/planningRuleConservation.test.ts — the suite, the REFRESH constant, and the constant-not-the-file assertion.
• motir-meta PR #238 / prompts/plan-rules/COMPRESSION.conserve.py — alignments(), as_addition(), the 11-case SELF_TEST and ALIGNMENT_GUARD to port.
• motir-meta prompts/plan-rules/COMPRESSION.md § Decision 1 — the chosen repair and the refusal of (B), including the mirror-obligation argument this card reuses.
• MOTIR-2969 — the motir-meta half; its comment thread carries the first measurement of this asymmetry.