Punctuation enforcement stays mechanical-only

Context

The writing-style guide staged its punctuation rules for a mask-aware docs-lint (itd-141): casing after colons and semicolons, and the em-dash list-item rule. The em-dash core shipped on 2026-08-28 as the punctuation/em-dash-in-list-item banned token (iss-2608280706531199), armed after the corpus was swept clean per the guide's warn-then-promote rule.

An adversarial design review then measured what remained against the corpus (docs/ + README.md, fence- and inline-code-masked):

• Colon-casing: 0 true positives, 3 false positives even in its narrowest scope (list-item pivots) — and all three are the project's own lowercase-by-design name opening a definition.
• Semicolon-casing: 0 true positives, 15 false positives — citation reference lists and acronyms, where a capital after a semicolon is correct English and mechanically indistinguishable from a violation without a dictionary.
• The em-dash residual (a wrapped continuation line of a list item): 0 instances.

The compounding fact: under the guide's promotion rule a check enters as a warn and is promoted to blocker only on a clean corpus. A check whose false-positive class cannot be closed is never corpus-clean at blocker severity, is therefore never promoted, and therefore never flips its row's label to machine-enforced — which was itd-141's entire goal. The goal is unreachable under the repository's own rules, not merely hard.

Decision

1. Punctuation enforcement is mechanical-only. What ships is the line-based banned token plus the <!-- docs-lint: allow --> escape. No dictionary-backed prose lint enters the native linter.
2. The colon and semicolon casing rules are review by nature, permanently. The guide records the reason on the rows themselves.
3. The continuation-line em-dash residual is accepted as review. Zero corpus yield does not warrant machinery (less-but-better; guards-prove-themselves). A first observed regression reopens it as a fresh intent (recurrence-is-signal).
4. itd-141 is superseded by this decision and retires to the supersession record.
5. Full-guide enforcement, if ever wanted, is an adoption, not a reimplementation: a Vale-class prose linter filed as its own intent, exactly as itd-141's SOTA section already ruled.

Alternatives Considered

1. Ship the casing lints with a lowercase-proper-noun allowlist. Rejected: an open-ended dictionary is the classic prose-lint failure mode, and no allowlist closes the semicolon class (any proper noun may follow one).
2. Ship the casing rules as permanent advisory warns. Rejected: warns gate nothing and never flip labels, so they add standing noise without moving the intent's goal (the spelling/* family is different — its warns catch real drift with a bounded escape set).
3. Build the continuation-line rule now. Rejected on yield: the corpus has zero instances, and the building blocks (fence mask, inline-span strip, list-state tracking) remain available the day a regression makes it worth a record.

Consequences

The guide's punctuation rows state final labels with their reasons; the "staged lints" forward-claim is removed; no row cites itd-141 as a staging promise. The supersession is recorded on both sides (this ADR's supersedes, the intent's superseded_by). The decomposition-calibration note carries the graded hand-run that led here.