Quarantine audio alignment per decision, not per whole call

Multi-turn audio reconstruction must score and quarantine each decision window independently after proving a recorder clock. A single per-call offset chosen by averaging all VAD spans makes one truncated or late lifecycle event invalidate otherwise usable earlier decisions, and the failure rate worsens with call length. First require authoritative sample-zero or sample-domain provenance; then validate each decision's causal bounds, caller-channel energy, and transcript-event link. Preserve failed decisions in audit receipts and exclude only their dependent examples. This does not rescue calls whose base clock is unauthenticated or justify lowering confidence thresholds.

Applies when: production voice calls are reconstructed into several decision-level audio training or evaluation examples. Keywords: LiveKit audio alignment VAD recorder clock multi-turn quarantine decision windows