Inference without a trail.

A clean endpoint for open-model inference: non-US infrastructure, retention off by default, and no training on requests.