fix(google_genai): attach grounding_metadata to streaming responses

[Abhijeet Prasad (AbhiPrasad)]

resolves #153

AI Summary

Issue

Google Search grounding metadata (grounding_chunks, grounding_supports, search_entry_point, web_search_queries) was silently dropped from span output when using streaming responses.

Root Cause

In py/src/braintrust/integrations/google_genai/tracing.py, the _aggregate_generate_content_chunks() function manually builds the aggregated output dict for streaming responses. It copied finish_reason and safety_ratings from the last candidate, but not grounding_metadata.

For non-streaming responses, the raw GenerateContentResponse Pydantic model was returned directly and serialized via model_dump(exclude_none=True), which naturally included grounding_metadata. So only the streaming path was broken.

Fix

Added extraction of grounding_metadata from the last chunk's candidate in _aggregate_generate_content_chunks(), serializing it via model_dump(exclude_none=True) to match the non-streaming output format:

Tests

Added 4 VCR-backed tests covering all transport modes with real Google API cassettes:

• test_google_search_grounding[sync] — non-streaming sync
• test_google_search_grounding[stream] — streaming sync
• test_google_search_grounding_async[async] — non-streaming async
• test_google_search_grounding_async[async_stream] — streaming async

Each test verifies that grounding_metadata on the first candidate contains web_search_queries, grounding_chunks, and grounding_supports.

Red/Green Evidence

• RED: Before the fix, stream and async_stream tests failed with AssertionError: Expected grounding_metadata on first candidate, got keys: ['content', 'finish_reason', 'safety_ratings']
• GREEN: After the 5-line fix, all 4 grounding tests pass, plus all 23 existing tests pass with no regressions