Python: [Bug]: Gemini streaming drops the model when a chunk has no model_version #9138

Description

GeminiChatClient._process_chunk passes chunk.model_version straight into ChatResponseUpdate.model, with no fallback. _process_generate_response, 33 lines above in the same file, falls back to self.model for the same field:

Both methods take the same type, types.GenerateContentResponse (:1034-1036 and :1066), and model_version is Optional[str] there. The SDK documents per-chunk absence for the very next field of that class, prompt_feedback: "Sent only in the first stream chunk."

What happens: ChatResponse.model is built only from the updates _types.py:2229-2230 keeps the last non-None value, and the Gemini client has no catch-up afterwards so when no chunk carries model_version, a streamed response ends up with model=None.

What I expected: the configured model, which is what the non-streaming path reports for the same input.

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

Steps to reproduce: drive get_response(..., stream=True) over a single chunk whose model_version is unset, then read updates[0].model. No API key and no network are needed.

Code Sample

from unittest.mock import MagicMock

chunk = MagicMock()
chunk.model_version = None


Feed that chunk through `GeminiChatClient.get_response(..., stream=True)` and read `updates[0].model`.

Expected: the configured model.
Actual: `None`.

Error Messages / Stack Traces

No exception is raised. The failure is silent: the field is simply absent
from the span attributes and from the metric dimensions

Package Versions

agent-framework-core: 1.20.0, agent-framework-gemini: 1.0.0b261002, google-genai: 2.27.0

Python Version

Python 3.11.4

Additional Context

MistralChatClient has the same asymmetry at _chat_client.py:800 / :830, but its SDK types CompletionChunk.model as required, so the gap is not reachable there. OllamaChatClient has no fallback on either path, so it is self-consistent. A PR for the Gemini case follows.