//! The VLM OCR companion: the in-process doc-VLM engine in its own executable,
//! so the default `anymd` binary stays small. `anymd setup ocr` installs it
//! next to the model weights and `anymd` runs it as the `__ocr-vlm-worker`
//! (page image path and token cap in argv, one JSON evidence document on
//! stdout), under the same supervision as before.