Fix merge_embed_responses dropping embedding types absent from first batch

[Sehastrajit-S]

Summary

merge_embed_responses() decided which embedding fields to merge by inspecting only the first EmbeddingsByTypeEmbedResponse in the list. If a later batch returned an embedding type (e.g. int8) that the first batch didn't include, those values were silently dropped from the merged result instead of being preserved.

Fixes #796.

Change

In src/cohere/utils.py, the field-selection check now looks across all batches (any(... for embedding_by_type in embeddings_by_type)) instead of only the first one, so a type present in any batch survives the merge.

Test plan

• Added test_merge_embeddings_by_type_preserves_type_absent_from_first_response in tests/test_embed_utils.py, mirroring the repro from Batched embeddings drop types absent from the first response #796.
• poetry run pytest (embedding utils tests + full non-network suite): all passing.
• poetry run ruff check / poetry run mypy clean on the touched files.

---

Note

Low Risk
Small, targeted change to batch merge logic for embeddings-by-type, with a regression test; affects embed client batching behavior when types differ across batches.

Overview
Fixes merge_embed_responses so embedding-by-type merges no longer drop types that appear only in later batches.

Field selection used to depend on the first response only; it now keeps any field that is non-None on at least one batch when building the merged embeddings_by_type result (e.g. int8 from a second batch is preserved even if the first batch had only float_).

Adds test_merge_embeddings_by_type_preserves_type_absent_from_first_response to lock in that behavior (issue #796).

^{Reviewed by Cursor Bugbot for commit bb499c7. Bugbot is set up for automated code reviews on this repo. Configure here.}