test(gate-acceptance): a sibling finding in the SAME file satisfies the subject too

[rubenvdlinde]

Revert-audits the last two acceptance bundles nobody had ever tested. #404 closed by asking for exactly this, on the grounds that #380's remedy — one mechanism per named file — is necessary but not sufficient.

debug-and-conflict — sound, unchanged

Four independent partial reverts, four predictions written before their runs, four matched on count and on the identity of the failing assertion. Dropping the whitespace tolerance from the call pattern reddens the Debugger.php row alone; narrowing the construct rule so it can no longer see a bare die; reddens the Terminator.php row alone. Those two lose disjoint findings, which is what proves the rows separable rather than merely both present. Exempting all Markdown reddens conflicted.md; reverting the setext discriminator reddens the clean arm instead — so both over-correction and regression are covered. No fixture change.

a11y-noise — was blind to the regression it exists to pin

FacesCamel.vue was authored to hold only the spellings the old word-boundary noun list could not see. But avatarUrl and avatar_url are two code paths, not one: camelCase splitting versus punctuation splitting. Reverting camelCase splitting alone left avatar_url firing, which kept the file in the log, which satisfied the fixed-string subject match — nine findings became five and the suite stayed green. Two more of the same shape were also green: dropping plain unbound src support, and dropping single-quoted tabindex.

Every full revert was caught. Full-revert-red is the test anybody actually runs, and it is why this survived three audits.

The invariant, one turn further

#404's defect was substring containment between subjects. This one is sibling findings inside one subject. The subject is a file path, and a file path is satisfied by every finding in that file. The construction that is safe by definition is a named file carrying exactly one finding: any mechanism that finding depends on empties the file when lost. Where one finding cannot state the claim, every finding in the file must ride the same code path — and "same code path" means reading the checker, not the fixture's comment.

a11y-noise goes from four rows to seven, with the corpus unchanged — the same nine and three findings, redistributed, so detection is bought without trading away a true positive. All three formerly-green probes now redden exactly one row each and name it, so which row fails tells you which mechanism was lost.

The planted files were renamed because Faces.vue is a prefix of FacesCamel.vue, and only the .vue extension separated the two subjects. Relying on an extension as the terminator is relying on an accident. Pairwise containment is now checked against the real log text, positive-controlled on #404's own corpus first, because a fixed-string search that cannot match reports "no containment" for free.

Two corrections

A sentence in debug-and-conflict/expect.conf claimed a regression would be caught "on the count". The driver never reads a count. This is the same wrong sentence #404 corrected in auth-guards — now found in two bundles written by two authors, so it is a recurring mental model rather than a typo. Fixed in place rather than deleted, because it was load-bearing for somebody's confidence.

COVERED-ELSEWHERE.md credits coverage by gate number while naming the providing suite in prose, and nothing ever resolved that name to a file — even though those rows feed the headline count of fixtured gates. All nineteen named suites exist today, so this check pins a true state rather than reporting a backlog, which is the cheapest moment to pin a registry. Built fail-closed: if fewer than ten names are extracted the check fails rather than reporting success over an empty list, because a check over zero names passes for free — which is the defect it exists to catch. Proven able to fail by renaming one credited suite.

Method

Thirteen probes, thirteen predictions written first, thirteen matched. Every revert quoted its actual changed line from the diff before running; Edit only, no scripted rewriting. The checkers and the runner were restored after every probe and are byte-identical to main — this branch touches only the acceptance fixtures, their expectation files, and the acceptance driver.

Assertions go from 121 to 131, none failing.

Findings: fleet-board/findings/bundle-audit-round2.md.