Compress natural language memory files into caveman-speak to reduce input tokens.

• Reduces token usage by removing articles, filler words, hedging, and redundant phrasing while preserving all code, URLs, file paths, and technical terms exactly
• Works on .md, .txt, .typ, .typst, .tex, and extensionless files; skips code files and mixed-content sections automatically
• Triggered via /caveman:compress <filepath> command or user request; creates human-readable backup as FILE.original.md before overwriting
• Includes validation and up to 2 automatic retries on compression errors; leaves original untouched if all retries fail