Round-3 cold state-machine review returned MERGEABLE (no fail-open, no permanent stall, all
five clause mutations pinned) with three LOWs, all measured rather than argued.
* The re-derivation `echo` asserted 'is NOT an attributable reviewer verdict' on a path also
reached by an ALLOW-LISTED creator whose row is in a state other than success/failure —
printing "creator='timothy', allow-list='timothy'". The ::warning:: below it was already
guarded against exactly that contradiction and the echo was not, which is the same defect
one line up. Now state-neutral.
* The failure short-circuit's message said an explicit rejection is left alone 'whoever
recorded it', but the guard requires attributability: a failure with a non-verdict
description, a null creator, or a DIFFERENT recorded base is still re-derived. Message
qualified rather than the guard widened — a bare [ state = failure ] would be safe by the
clause's own argument, but it would quietly opt `failure` out of the base binding, and
none of the excluded shapes is producible by post-review-verdict.sh. Written down so the
next reader does not 'simplify' it.
* ci.exemption-provenance cited a test that the success/failure split had renamed — a
mutation proof nobody could run. Fixed, and the class is now guarded: a new test derives
the cited names from the record's own text and asserts each is defined in scripts/tests.
Mutation-proven against the pre-fix wording.
scripts/tests: 970 passed, 2 skipped.
fixes#742
Decisions-Edit: yes