Project Layout

The package is organized by RF domain ownership. waveforms, domains, engine, receiver, placement, scene, observation, labels, and storage implement the local generation and persistence path. annotators and inference implement prompt projection, templates, structured output, and the provider-neutral extension boundary. annotation owns the unified Spark annotation runner and the Dataproc bootstrap; the local HDF5 overlay, the immutable remote request/sidecar publication, and Vertex Batch control went with the WebDataset annotation route. There is no golden_paths package: the Golden Path contracts are tests, at tests/golden/, and reusable generation and storage logic lives in its domain packages.

Within rfgen.scene, composer owns scene realization and rf_context owns receiver-frame resolution. Configuration lives in rfgen.config; command composition lives in rfgen.cli.app, while generation execution and its command live in rfgen.generation. Tests and the two production-qualified Golden Paths validate the installed package boundary. The authoritative directory taxonomy and execution policy are in the test map: live provider/cloud tests are top-level tests/manual/, and real Sionna RT journeys are tests/integration/sionna_rt/. A tests/qualification/remote_annotation/ directory stood here; it went with the remote WebDataset annotation route it qualified.

The repository contains a bounded Dataproc Serverless submission path that publishes an immutable GCS Signal Dataset snapshot. It is available only for manual validation; it is not production-qualified. The narrowband classifier and chirp-radar Golden Paths remain the production-qualified execution journeys. The repository contains no general managed-cloud deployment platform or mutable remote annotation store. Its remote annotation capability is a Dataproc Serverless batch that publishes a dense annotation set under the snapshot root; the Vertex Batch workflow over WebDataset shards that stood here was retired with that store.