[jserv]

Interviewer transcription describes generated words, not what the candidate heard. When queued audio is cut by an interruption, a stale-playout drop, a thinking hold, a pause or a socket handoff, the transcript used to keep those words as ordinary dialogue, so recovery, interim review, the phase judge and the report could assume a question or hint was delivered. This records delivery uncertainty beside the unchanged raw transcript, keyed by interviewer line, and every model-facing reader substitutes a marker for the whole marked turn. Raw text stays for independence checks, grounding, search and byte paging, and conversation lookup flags affected rows and continuations.

The rules sit in InterviewerDelivery so every cut path shares them. A cut of a running generation marks its later fragments until its completion. A cut of completed or stalled output marks fragments until the next audio, because transcription trails audio and that completion may never come. A cut that stops nothing marks nothing. A cut seals the line it marks, so the next reply opens its own line after the candidate's words instead of joining omitted text. Words a reply sends ahead of its own first audio after such a cut cannot be told from the cut tail and stay marked; nothing after that audio is. The policy is in docs/interviewer-delivery-evidence.md, and bundle 39 carries live prompt 30 and report prompt 23.

Verified with ./scripts/test.sh (all lanes, no reported skips). Each marking, window, sealing and continuation rule was broken by hand and a test failed for each. Not covered: credentialed media probes and real provider event ordering, which the conversational media benchmark measures separately.