Create Predict

Create and execute a prediction-only job. Distinct from /pipeline: runs prediction (and the minimum upstream prep, if needed), but exposes a separate session lifecycle and result endpoints. Use /pipeline for the full classify/apply flow.

Two input modes (auto-detected at job start):

• Mode A — dataset already prepared. Triggered when dataset_path points at a readable dataset.jsonl, or when an existing session_id has cache/dataset/dataset.jsonl whose referenced images resolve. Only the predict step runs.

• Mode B — USD upload / s3_uri / fresh session_id. When no runnable session dataset is present, /predict runs the minimum upstream steps (optimize_usd if enabled → identify_asset → build_dataset_usd → build_dataset_prepare_dataset) before predicting. apply_physics is intentionally not part of /predict — use POST /pipeline if you need the full classify/apply flow.

  A cached JSONL whose images do not resolve falls back to Mode B when an input USD is available. Without an input USD, the request returns HTTP 400 with recovery guidance.

The detected mode is persisted to session metadata under predict_mode and surfaced in the GET /predict/{id}/results response.