docs: optimize AI agent instructions for token efficiency

- Reduce copilot-instructions.md from 416 to 77 lines (~95% reduction)
- Create AGENTS.md with operational procedures (on-demand loading)
- Implement thin pointer architecture to minimize auto-loaded tokens
- Auto-loaded context reduced from ~10,000 to ~500 tokens per chat

Instructions now use tiered loading: essential context auto-loads,
detailed procedures load on-demand via #file:AGENTS.md reference.

Author: nanotaboada