Add gRPC service for the activity classifier postprocessor

[joshhvulcan]

Motivation

The activity classifier's postprocessing rules (transiting, anchored/moored, stationary, confidence thresholding, geofencing, near-shore/marine-infra, etc.) currently only run in-process inside the atlas_activity model pipeline. We already expose the entity postprocessor as a standalone gRPC service (atlas_entity/grpc_serve) so it can be deployed and invoked independently of the model. This adds the equivalent for the activity postprocessor, so the same rule logic can be deployed and run separately from the model pipeline.

Changes

Adds a self-contained atlas_activity/grpc_serve package mirroring the entity one: a .proto + checked-in stubs, config.py (ACTIVITY_POSTPROCESSOR_GRPC_*, default port 5208), a serve_grpc.py server with a health check, an integration test, and a test_activity_postprocessor_grpc docker-compose service.

Unlike the flat entity input, the activity postprocessor needs the trajectory itself, so the request carries the AIS messages (sog, rel_cog, cog, lat, lon, send, optional nav/dist2coast) plus the predicted class, model confidence, and binned_ship_type. The servicer reconstructs the most_recent_data DataFrame and calls the existing AtlasActivityPostProcessor.postprocess — no rule logic is reimplemented, so behavior is identical to the in-process pipeline.

Architecture

flowchart LR
    client[gRPC client] -->|ActivityPostprocessorInput| svc
    subgraph atlas_activity
      svc[ActivityPostprocessorServicer]:::added -->|rebuild DataFrame| pp[AtlasActivityPostProcessor.postprocess]
      pp --> rules[postprocess rules]
    end
    svc -->|ActivityPostprocessorOutput| client
    classDef added fill:#6f6,stroke:#333

Loading