Playback evidence must drive the same state that chooses recovery speech

Playback evidence must drive the same state that chooses recovery speech. A voice system can correctly measure acknowledged audio yet still omit required information if a separate recovery rule assumes that information was delivered. Track short information units against their playback receipts, and use those receipts for both conversation context and replay decisions. Persist the measured endpoints at interruption; final counts alone cannot reconstruct a cutoff. This applies to streamed voice systems with reliable receipt-to-audio association. Provider playback is not proof of human hearing, and uncertain units may need repetition.