Start both roles at once, and stop the interview deadlocking

compute_diff now seeds the standing questions on the run when it opens, so the author starts answering from the session while the reviewer is still paging the diff. Across 15 recorded runs that phase was 670 seconds of a 1,400-second median, and it was serial only because the author had nothing to answer until the reviewer handed back question ids.

A cap that only lived in prose

The interview is capped at two batches of questions, but only the role prompts said so. Three recorded runs broke it the same way: the reviewer opened a third batch after the author had correctly handed back, nothing was left to answer it, and the reviewer polled get_answers for 61 to 103 minutes. On one run that was 89% of the wall clock.

record_interview_round now refuses a third batch that carries new questions, and points the findings where the protocol already sends them: verification.not_verified and the reviewer’s closing report. Re-recording a question the run already holds is still a free retry.

Telling a waiting reviewer the truth

get_answers ended with “the author has not replied to those yet”, which reads as “keep waiting”. The author is a subagent and does not idle: if it has stopped, it has returned. After 8 polls or 60 seconds with no new answer, the tool now says so and names the two real options, re-dispatching the author or submitting with these as findings.

Also

• •Every submit failed again. summarizeRun grew a field, submit_document spread the whole summary into meta.interview, and that block allows exactly five. The stamp is now narrowed to the schema’s own fields, so the next field added is a compile error instead of a failed distillation.
• •The distillation diagram was redrawn. It had drifted past the house-rules tool, and it drew a private channel between author and reviewer that does not exist: the interview goes through the server precisely so it can attest who answered.