Python: Honor the configured model in streamed Gemini chunks

[Ayoub (Ayoubhm07)]

Motivation & Context

_process_chunk passes chunk.model_version straight into ChatResponseUpdate.model, while _process_generate_response falls back to self.model for the same field 33 lines above. _types.py:2229-2230 keeps the last non-None update value and the Gemini client has no catch-up afterwards, so a stream whose chunks carry no model_version ends up with ChatResponse.model is None where the non-streaming path reports the model.

observability.py:3551-3552 sets gen_ai.response.model from that field, and RESPONSE_MODEL is in GEN_AI_METRIC_ATTRIBUTES (:3558): _filter_metric_attributes drops absent keys before _capture_response records the token and duration histograms (:3604-3610). On Gemini streams those lose their per-model dimension, and _backfill_request_model (:1905) cannot recover the span name either.

Description & Review Guide

• What are the major changes? One line, applying the expression already on :1058 to :1091. Both methods take the same types.GenerateContentResponse (:1034-1036, :1066), and model_version is Optional[str] in the pinned google-genai 2.27.0 the SDK documents per-chunk absence for the very next field of that class, prompt_feedback: "Sent only in the first stream chunk."
• What is the impact of these changes? A chunk that carries model_version is untouched. Only the update's model changes, and through aggregation ChatResponse.model. Two tests added next to the existing non-streaming one; only the fallback test fails on main, the other guards the pass-through and passes already. pytest packages/gemini/tests 254 passed 10 skipped, ruff check and ruff format --check clean, pyright 0 errors on the changed source, mypy packages/gemini/tests clean all at the versions in uv.lock.
• What do you want reviewers to focus on? Whether self.model is the right value here: it is the client default, not the model resolved from options at :549, which _process_chunk does not receive. That matches :1058 exactly, so I kept parity rather than changing both happy to thread the resolved model into the two parsers instead if you prefer. MistralChatClient has the same asymmetry at :800 / :830, though its SDK types CompletionChunk.model as required so the gap is not reachable there; happy to follow up separately. ollama has no fallback on either path, so it is self-consistent and out of scope.

Related Issue

Fixes #9138

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR is linked to an issue and there is no other open PR for this issue (see Related Issue above).
• This is not a breaking change.

[Copilot]

Copilot review overview

🟡 Changes recommended

The fallback mislabels responses when request options override the client’s default model.

Review effort: Balanced
Findings: 1

Open (1)

• Use resolved request model when labeling streamed responses · New

What changed in this PR

Adds Gemini streaming model fallback to preserve response metadata and observability dimensions.

Changes:

• Falls back to the client model when streamed chunks omit model_version.
• Adds tests for pass-through and fallback behavior.

File	Description
_chat_client.py	Adds streamed model fallback.
test_gemini_client.py	Tests streamed model handling.

---

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

[Ayoub (Ayoubhm07)]

_prepare_request already resolves the model and it was in scope at the call site so _process_chunk now takes it and uses it as the fallback instead of self.model

I added test_get_response_streaming_model_falls_back_to_requested_model: a gemini-2.5-flash client called with options={"model": "gemini-2.5-pro"} over a chunk with no model_version now reports gemini-2.5-pro.

_process_generate_response has the same client-default fallback at :1058. I left it alone to keep this PR on the streaming path happy to align it here or in a follow-up, whichever you prefer

[Ayoub (Ayoubhm07)]

Resolved the conflict with main. The only overlap was the _process_chunk call, which 7a262d6a moved into the new try/finally that closes the SDK stream; the call now passes model from inside that block. The diff against main is unchanged otherwise: 4 lines in _chat_client.py and the 48 test lines.