total_token_details sums per-category token counts across traces in the
thread. Keys are model-specific category names (for example cache_read,
cache_write, reasoning, audio).
Example: {"cache_read": 400, "reasoning": 120}.
total_token_details: __type