An image sent to /v1/embeddings for a model with no image encoder ran as its
one placeholder token, and a text embedder such as EmbeddingGemma 2 answered
200 with 768 zeros. add_request now refuses an image input when the
pipeline's input modalities lack Vision, before any sequence exists.
/v1/embeddings folded every engine ValidationError into a 500; it now answers
them 422, as it already does its own checks and as the other endpoints map
validation errors.
Closes#122
Co-authored-by: Claude Opus 5.5 <[email protected]>