fix(writeback): a REVERT closes nothing — refuse the quoted title and the wrapped ref (closes card#8306) - #620

GitHub composes a revert PR by quoting the original's title and wrapping its branch. Once card#8294 made (closes card#N) the title convention, every revert inherited a closing form — and a card token — for work it UNDOES, so merging one moved the reverted card to Shipped and the board asserted the opposite of the truth.

What's in it

• app/Bridge/Support/RevertGrammar.php (new) — the one authority for the two artifacts GitHub mints for a revert: withoutQuotedRevert(), isRevertRef(), isRevert(), describeRefusal().
• ClosureGrammar::remainders() subtracts the quoted span at the single private choke point all four public predicates read tokens through.
• PrOutcome::mergeClosesCard() gains a ! RevertGrammar::isRevert($title, $headRef) conjunct — both surfaces — and takes the title as a required 4th argument so no caller can silently keep the ref-only reading. The structural route's own premise (the ref is the CARD's branch, carrying its work) fails on a revert, whichever surface says so.
• The withheld-merge warning and bridge:reconcile's skip line gain a revert arm. Both asserted the ref does not name the card and the title carries no closing form — untrue on both counts for a revert, which would have sent an operator to rewrite correct prose. Every non-revert line is byte-identical.
• pr-title-lint.yml's "⚠ DISCLOSED, not fixed here" comment is discharged in place. The revert-* exemption stays — it governs what CI DEMANDS, never what the writeback READS.
• DL-318, docs/writeback.md § Mention vs closure, docs/config-schema.md, docs/kanban-integration-contract.md, CLAUDE_CONVENTIONS.md § PR titles, docs/CHANGELOG.md.

Both routes were live — measured, not reasoned

End-to-end through the shipped classifier over a merged-revert payload before any code changed:

⚠ The last two rows were found by adversarial review of the first revision of this PR, not by me. That revision gated the structural route on the REF wrapper alone, so a hand-made git revert — which wraps no ref and announces itself only in the title — still promoted the card it reverted, and did so silently: a closing card never reaches warnMentionWithoutClosure(). Its PrOutcomeTest pinned that behaviour as intended. The conjunct now asks isRevert() on both surfaces and the assertion is flipped, with the rows above as witnesses at the predicate, end-to-end in classifier witness 9b, and at the bridge:reconcile backstop.

The original structural hit is CardTokenGrammar::parse('revert-611-card-8294-slug') === 8294 — GitHub's wrapper carries a copy of the original branch's name, so a title-only repair would have left half the hazard open on every install still on the card-<id>-slug convention the toolkit mints.

The payload shapes are OBSERVED

No revert PR exists on any repo this shop owns (1,607 PRs scanned, zero on title or head ref), and GitHub's Reverting a pull request article states neither format. Both were pulled through the REST API from public repositories where GitHub minted them: laravel/framework#61330 (Revert "<the original's title>" on revert-61320-fix/contains-strict-comparison, against #61320 whose head is fix/contains-strict-comparison) and #61262 (revert-61245 — the tail is dropped when the original's branch is gone).

Gotchas

• ⚠ CHANGES WHAT THE WRITEBACK ACTS ON. No migration, no config key, no token-scope change, no receiver accept/reject change, and the pr-title-lint accept-set is byte-untouched.
• ⛔ Correlation is untouched. A revert still stamps PR refs and the ungated outcomes still fire exactly as DL-305 §2 ruled; only the completion claim is refused.
• ⛔ A revert of a revert does not close either — GitHub does not escape the inner quotes so the depth is unparseable, and the first revert never moved the card back, so the re-apply has nothing to promote.
• ⭐ To close a card ON a revert, put the closing form OUTSIDE the quoted title. Positional, no new vocabulary; [no-close] stays the CI-only inverse it was. Its reach is bounded by correlation — it re-closes the card the branch names and cannot redirect the move.
• ⚠ A divergence stays open, and DL-318 IS its filing (no separate card, stated so deliberately): the lint decides is this a revert from the revert-* BRANCH while the runtime also reads the TITLE, so a hand-made git revert on an ordinary branch passes CI and is refused at runtime. Under-promotion, loud, recoverable — the same shape already recorded for the branch predicate.
• ⚠ THE PRICE OF READING THE TITLE, priced rather than asserted away: a PR whose title merely QUOTES the word (fix: revert "the streaming change" and re-land it) now loses its structural route. It is the identical over-refusal the lexical route already accepts by construction, it fails toward under-promotion, and it is LOUD — the card lands in the withheld set so the warning fires and names the revert. 0 of the 1,566 merged PR titles across all three repos match it, measured here with a known positive planted into that same corpus to prove the grep discriminates. (The first run of that grep had a broken control and its 0 meant nothing; the figure above is the re-run.)
• ⚠ opened fires but usually does not MOVE. A revert PR still emits the opened target; KanbanMoveCardHandler::isRegressiveMove() makes opened strictly forward-only, so in the dominant case — a revert of a merged PR whose card reached Shipped — the move is refused. It moves only where the card still sits behind In-Review.
• ⚠ The obvious DL-239 move was tried and withdrawn: adding a revert row to ClosureGrammar::VECTORS reds two PrTitleLintTest ties (the corpus is title-only while the lint's revert handling is branch-keyed; and the message tie derives a verb stem from each accepted row's first word). Both constraints were undocumented until this card measured them and are now in the VECTORS docblock, pinned by a test.

Verification

• 13 mutations over the new predicate, all RED, each proven to have landed (sha256 before/after + the mutated text read back) and restored byte-identically: lexical subtraction deleted · structural conjunct deleted · structural conjunct regressed to the ref-only reading · structural conjunct reduced to the title surface only · reconcile passes a blank title · classifier passes a blank title · unterminated-wrapper alternative dropped · quoted span made non-greedy · ref anchor dropped · ref digits dropped · classifier warning arm removed · reconcile skip arm removed · isRevert reduced to one surface.
• 2730/2730 phpunit (10312 assertions, SQLite) · phpstan L7 0 · pint clean · check-doc-refs clean · decision-log.py check clean.