Start Batch Predictions Job

Start an async batch prediction job.

Two input modes (mutually exclusive):

• examples — ingest brand-new examples and run predictions on them.
• example_ids — re-predict existing examples (referenced by their ClickHouse ids). Predictions are inserted as new versions on the existing example rows; no new examples rows are written.

When compute_confidence=True the worker pre-warms a probe scorer container per arbiter and, after predictions persist, kicks off Modal batch confidence scoring filtered to the prediction_ids produced by this job (so we don't accidentally re-score unrelated NULL-confidence rows). Predictions are visible to clients during the scoring phase via GET /results — confidence populates as Modal streams.

Phases visible via GET /{job_id} and SSE: predicting → scoring → done (or failed).