Race Cameras: Timing, evidence and recovery

Timing, evidence and recovery

Synchronize the station clock with the timing machine before racing. All live camera observations remain live_arrival_estimate; network buffers and decoding/OCR introduce delay. Calibrate the timing offset against a reference clock/video. SRT receive buffering is separate from this offset. Use a small, legible crossing/bib ROI so PNG evidence remains below 256 KB. Check the preview, frame age and detection confidence before accepting crossings.

Each capture session runs for at most 10,000 sampled frames. Keep the station open; closing it stops capture/replay processes. Closing Central Control does not stop the station. During upload failure, observations are saved in station-local JSONL logs and capture continues after bounded upload retries. Upload retries can delay sampling, so this is not a zero-loss/real-time timing guarantee. Re-pair/reconnect when needed, then use Replay logged detections or Saved station detection logs → Replay selected log. Saved archives remain available after station restart. Replay retains original camera IDs/timestamps, and Central Control deduplicates identical observations. Logs/previews are stored beside the station database in <database-stem>_cameras; preserve them with race evidence. Replay jobs report running/completed/failed in station status.

Central Control grants this station an upload-only token for its camera namespace, not the central account password or control rights. The station pairing key authorizes station control. Camera IDs are prefixed with the station ID; shorten the camera ID if the combined ID exceeds 64 characters. Pairing/credentials are not returned in session status.

Verification: 45 Python tests passed. A two-process localhost integration test paired Central Control with a separate station, remotely started an RTMP listener, received video/PNG preview and OCR observations through scoped upload authentication, disconnected it, and replayed without duplicates. Station restart/re-pair and saved-archive replay also passed. This simulates separate machines using separate processes; cross-machine TLS and real NDI hardware were not tested.