RTK compression

Compresses tool_result blocks (long file dumps, search results, terminal output) into compact summaries before dispatch. Typical savings: 20–40% input tokens per request. Fail-open: if the compressor errors, the original payload passes through untouched.