[austinjiann]

A pending operator question could still allow a stale agent to launch a stage, and a reply arriving as a turn ended could be swallowed by the supervisor cursor until its timer expired. Hold checks now run under the launch lock for stages and agent processes, and the supervisor retains the journal cursor the session actually observed.

The change also preserves transcripts and session IDs when a question interrupts a watchdog resume, clears genuinely stalled sessions, and adds --run-id RUN_ID --watch for a scoped worker that stays available for replies. Existing named one-shot behavior is preserved.

Validation:

• 1,039 Python tests pass; repository formatting and Python lint pass.
• Independent code review completed; two edge findings fixed and re-reviewed.
• Real Codex CLI worker + local admin: asked a question, received the browser reply, resumed the same session, and recorded the token in 9.1 seconds with a one-hour quiet timer. Both saved bottle hashes matched before and after.
• Zero stage executions or provider commands; paid ledger unchanged. Diagnostic paused and supervisor stopped. This verifies live reply/resume, not fresh reconstruction or quality acceptance.

Local evidence: .context/evidence/worker-reply-resume/. The system CLI was separately repaired from broken 0.130.0 to official 0.163.0; that machine setup change is not part of this diff.