A reviewer row now says in one sentence who the reviewer is (the frozen
route display name), what it was asked for at start, and at settlement
whether it answered, how it ran, or why it did not (the engine's reported
sentence, quoted; a spent window's reset instant); a zero-cost refusal
reads as not sent. No code, slot id or custody state reaches the row.
The wave's verdict line is one family for every collected wave, k of n
reviewers answered with the findings named plainly; the DEGRADED token,
the paid-cycle tail, the slot reasons and the late-result tail leave the
line and stay in Reviews and the task detail. The open-wave lines, the
no-dispatch line and the sibling lines of the same family lose their
internal words.
The rationale the plan author already writes is emitted as a narration
row in the model's own voice at both author paths; an unbound ToolContext
accepts the narration keyword. The three model-facing strings that promised
a loud host disclosure now say the open review stays typed in the task's
state and the answer is where it is said in words. The mind's own rendering
of a failed reviewer shows the reported cause instead of the engine's retry
advice, and the custody paragraph counts the reviewers that did not answer.
A delegated review session that fails with no engine code was recorded only as
the host-derived state code run_failed; the sentence the engine reported
(failure.safeMessage) survived only inside the error prose. The gateway now
translates a failed run into the same typed refusal as before and carries the
reported words as one opaque, redacted, bounded fact, reported_cause, which the
reviewer record, the shared typed-failure key list, the frozen triad/scope
roster and the durable triad row keep. Nothing branches on the text.
