A vendor serves /v1/embeddings, held to what its model gives

Vendors can serve /v1/embeddings for the OpenAI dialect, and routing rejects chat models for embeddings and the reverse.