A spatial observer and the first voice file

Spatial work started by testing normalized geometry, zone boundaries, and hysteresis. Replay then showed that tiny detections and multiple detections in one frame could produce misleading zone changes. The observer was revised to choose one qualified primary target per frame, using area first and confidence to break ties, while retaining secondary targets.

The revised Phase 4D replay processed 99,669 frames and recorded eight zone transitions. A surrounding-frame audit supported the remaining changes. The observer then became the sole live spatial observer, producing telemetry while production perception remained healthy. It stayed advisory-only: it had no authority over behavior or actuators.

The production perception gate was preserved, but the launcher had been changed earlier to add a raw-output tap, with one intentional service restart. The later observer handoff did not restart production perception. This distinction matters when describing what was and was not changed.

Phase 5 continuity experiments stayed offline. The revised strike-and-recovery logic improved agreement with the reference replay, but a parameter sweep did not justify promoting a winner. Agreement with largest-area selection was a comparison measure, not proof of physical identity.

Piper generated a 4.725-second voice WAV at 22,050 Hz, mono, 16-bit PCM. File synthesis passed. No speaker was connected, so audible playback did not receive a pass.

The reSpeaker XVF3800 assembly was received and identified through its USB and DFU interfaces. Firmware and documentation were located and checked. The work stopped before any firmware write, and the module was disconnected. USB runtime capture, direction-of-arrival behavior, and physical playback remained unqualified at this checkpoint.