Add an explicit source delivery policy to the detection pipeline so offline and realtime sources can be handled differently without splitting the runner.
Blocking sources now backpressure ingestion until their pending frame is drained, which preserves every offline video frame even when inference is slower than decode. Latest-only sources keep the previous overwrite behavior for realtime feeds such as cvmmap.
The tests now cover both policies: offline sources retain ordered frame delivery under slow inference, while latest-only sources still drop intermediate frames as intended.
Refactor the package into common, schema, detection, and tracking namespaces and move dataset-specific ActualTest utilities into tests/support.
Add a pluggable detection stack with typed protocols, pydantic-settings config, loguru-based runner logging, cvmmap and headless video sources, NATS and parquet sinks, and a structured coco-wholebody133 payload path.
Teach tracking replay loading to consume parquet detection directories directly, preserve empty frames, and keep the video-to-parquet-to-tracking workflow usable for offline E2E runs.
Vendor the local mmcv and xtcocotools wheels under Git LFS, update uv sources/lock state, and refresh the mmcv build so mmcv.ops loads successfully with the current torch+cu130 environment.
crosstyan