_To_protocol_usage Drops Input_token_details In V3 Streaming Path, Causing Cache Token Counts To Always Be 0

In langchain_core/language_models/_compat_bridge.py, the helper _to_protocol_usage builds the protocol usage dictionary by explicitly selecting only input_tokens, output_tokens, total_tokens, and cached_tokens from the model's usage_metadata. It silently discards the nested input_token_details and output_token_details dictionaries, so the final aggregated protocol event never carries cache_read or cache_creation counts.

Run the provided reproduction code that simulates a streaming chat model yielding a final ChatGenerationChunk with usage_metadata containing input_token_details. Collect events via achunks_to_events and inspect the message-finish event. The usage dict lacks input_token_details even though it was present on the chunk.

This change widens the protocol usage payload. Downstream consumers or strict validators that do not expect the nested input_token_details/output_token_details keys may fail or ignore them; ensure the protocol schema and any deserialization models accept these fields. The conditional checks use is not None to avoid adding empty objects, but if usage_metadata contains other unknown nested keys they are still dropped. Rollback by reverting to the previous explicit field selection if downstream breakage is observed. No thread-safety impact because the function only reads from a local dict, but concurrent stream aggregation must ensure the final chunk with details is not overwritten by later chunks lacking details.