Fix KeyError when parsing video metadata without audio track in Google models

[jerry-heygen]

This PR fixes #2489.

Problem

When processing video content that lacks an audio track through Google's Gemini models, the metadata parsing fails with a KeyError. The API response includes an AUDIO modality entry in the metadata details, but without the token_count field, causing the code to crash when trying to access it.

This can be reproduced by adding the following line in pydantic_ai_slim/pydantic_ai/models/google.py#L606:

for key, metadata_details in metadata.items():
    if key.endswith('_details') and metadata_details:
        suffix = key.removesuffix('_details')
        for detail in metadata_details:
            details[f'{detail["modality"].lower()}_{suffix}'] = detail.get('token_count', 0)
            print("DEBUG: detail: ", detail)  # <-- Add this debug line

processing a video file without an audio track. Debug output shows:

DEBUG: detail:  {'modality': <MediaModality.TEXT: 'TEXT'>, 'token_count': 1201}
DEBUG: detail:  {'modality': <MediaModality.VIDEO: 'VIDEO'>, 'token_count': 2064}
DEBUG: detail:  {'modality': <MediaModality.TEXT: 'TEXT'>, 'token_count': 1717}
DEBUG: detail:  {'modality': <MediaModality.AUDIO: 'AUDIO'>}

Solution

Use .get('token_count', 0) instead of direct key access to safely handle missing token_count fields:

# Before (causes KeyError):
details[f'{detail["modality"].lower()}_{suffix}'] = detail['token_count']

# After (safe with default):
details[f'{detail["modality"].lower()}_{suffix}'] = detail.get('token_count', 0)

This ensures that modalities without token counts (like audio tracks in silent videos) default to 0 tokens rather than crashing the application.

Testing

• Verified the fix resolves the issue with video files lacking audio tracks
• All existing Google model tests pass without regression