Problem
Current large query results are sent entirely to LLMs, causing:

• Token waste
• Performance issues
• No reusable outputs for downstream workflows

Solution
Add export_to_json tool that:

1. Streams results directly to filesystem
2. Returns only minimal metadata to LLM:

   {
    "success": true,
    "file_path": "/exports/query_12345.json",
    "count": 15000,
    "sample": [first_3_docs]
   }

3. Preserves data for:
   • Visualization tools
   • Subsequent MCP processing
   • Human verification