Cohere Embed v4 returns one vector for several images when input_type is image, against the documented one-vector-per-input promise

[JGoutin]

Observed while working on #126/#193, in code that shipped before this milestone.

On cohere.embed-v4, sending N images with input_type="image" returns one vector, while the same N images with input_type="search_document" return N. The gateway forwards input_type faithfully, so this may well match Cohere upstream — but docs/api_cohere_embed.md promises one vector per input without qualification, and that promise does not hold in this combination.

Two things to settle:

• Whether Bedrock/Cohere really behave this way for that parameter pair, or whether the gateway is mis-shaping the request. Verify against the Cohere reference and a live call before changing anything.
• If it is upstream behaviour, the page must say so — a documented limitation is acceptable here, an unqualified promise that is false is not.

Ref: docs/api_cohere_embed.md, stdapi/models/embedding/cohere_embed.py, https://docs.cohere.com/reference/embed