Round-3 cold tests/prose review returned BLOCKED. Its findings were all of one shape: the
clauses are correct, but the suite could not tell.
Measured green with the mechanism DELETED, now each reddens on exactly one test:
* deleting the ::warning:: entirely — 166 green. It is not decoration: docs/ci-cd.md heads a
paragraph with 'If a reviewer's verdicts stop being honoured, this is why', and
ci.exemption-provenance uses its existence to soften the #845 residual. Asserted present
in the off-list case, and ABSENT for an allow-listed reviewer, which also pins the
ex_human != yes conjunct.
* dropping any ONE clause of the changed-row triple, or reducing the triple to a single
comparison. Both mid-run tests moved all three fields at once, so 'compare the whole row'
was argued, not measured. STUB_MIDRUN_ROW plus one parametrised case per field.
* a membership loop that can only ever match its LAST entry. Every behavioural test ran the
singleton list, under which that is indistinguishable from a correct loop — and the
annotation this change adds tells an operator to ADD a second entry, so the two-entry path
is the first one exercised in anger. A reviewers= override substitutes a two-entry literal
into the real step body; first-entry, last-entry control and absent-account control.
Prose, all measured by the reviewer rather than argued:
* 'was inherited as a human verdict' described an event nobody observed — renovate has never
posted a commit status in this repo (every row on all six Renovate-authored PR heads carries
creator: null). The asymmetry underneath IS measured; the ROUTE is an inference. Conditional
voice in all three places, with the distinction stated.
* the ::warning:: does not fire for a row recording a DIFFERENT base — the base binding
re-derives it before the annotation is reached. 'whenever' was false.
* a credential drift re-derives to an exemption success on an exempt PR, not only to pending —
the quieter and worse half.
* the live account-set measurement now carries its two curl paths.
* CLAUDE.md states the H10 convention and now carries the constraint: the credential you post
with must be an allow-listed account, or the verdict is written, reported as success, and
silently re-derived (#845).
* residual session chronology trimmed from the record.
STUB_PRE_ROW now emits the measured shape (total_count, a decoy context) rather than a
minimal one.
scripts/tests: 976 passed, 2 skipped.
fixes#742
Decisions-Edit: yes