Reviewer subprocess failure leaves no diagnostic, consumes a review number, and advances the spiral round count

[HenryLach]

Report

penster batch 20260911T234647, field report on #627, item E. Twice (different lanes) the first Step-5 code review's reviewer subprocess died silently: .reviewer-state.json left at status=running, 0 tool calls, no verdict file content, no reviewer events or stderr written. The runtime's retry ~1 min later succeeded. Side effects:

1. The dead attempt consumed a review number (R003), so the retry's verdict landed as R004.
2. The spiral detector's "round N" label counts spawns, not verdicts — it reported "round 4" for the second real verdict, and its consecutive-REVISE threshold is therefore reachable one attempt early.
3. Nothing on disk explains the failure (no <reviewer>-exit.json, no stderr tail).

Expected

• A reviewer spawn that produces no verdict writes a diagnostic: the agent-host exit summary + stderr tail under .pi/runtime/<batch>/agents/, and a Review spawn failed line in STATUS.md's execution log with the cause (exit code / signal / timeout).
• The review number is allocated on verdict write, or the failed attempt's number is reclaimed (no gap, no phantom round).
• review-analysis round counting keys on verdict-bearing review files for the gate, not on review_started events; a spawn failure does not advance the round or the spiral streak.

Acceptance

• Behavioural test: mocked reviewer spawn that exits without writing a verdict → diagnostic files present, STATUS line present, next successful review is R00N (same number), boundary notifications label it round 1.