ML deps (torch, transformers, onnxruntime, opencv) are ~4GB and stay out
of the regular web/worker image. Models self-heal into /models on start
(implemented in FC-2). HuggingFace cache vars point inside /models so
weight downloads are persistent across restarts.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
