Module batch 

Shared batch processing trait and entry types.

Design intent

Most perception stages are per-feed: each feed has its own detector, tracker, classifier, etc. However, inference-heavy stages (object detection, embedding extraction, scene classification) benefit from batching frames across multiple feeds into a single accelerator call.

The BatchProcessor trait captures this pattern. The runtime collects frames from multiple feeds, dispatches them as a batch, and routes per-item results back to each feed’s pipeline continuation.

Backend independence

BatchProcessor does not assume ONNX, TensorRT, OpenVINO, or any specific inference framework. Implementors choose their own backend.

Temporal window support

Each BatchEntry carries a single FrameEnvelope — the right granularity for single-frame inference.

For models that operate on temporal windows or clips (e.g., video transformers, clip-based action recognition), the extension path is:

1. A per-feed stage assembles the frame window from internal state or the temporal store.
2. The assembled window is stored as a typed artifact in StageOutput::artifacts for a downstream stage, or the batch processor manages its own per-feed window buffers internally.

This keeps the BatchEntry type and coordinator protocol focused on the single-frame case.

Structs

BatchEntry

An entry in a batch, passed to BatchProcessor::process.

Traits

BatchProcessor

User-implementable trait for shared batch inference.