Use json.dumps with ensure_ascii=False to avoid converting Chinese
descriptions into Unicode escape sequences.
This makes logs easier to read and significantly improves LLM tool
selection accuracy (from ~48% to ~67%) by preserving the original
natural language context.