chore: update CLAUDE.md with agent protocols

[jdalton]

Summary

Adds structured agent behavior protocols to CLAUDE.md:

• Pre-Work: clean dead code before refactoring, phase multi-file changes (≤5 files), study references before building, work from raw error data, execute immediately on "yes"/"go"
• Verification: run actual commands, state what was verified, forbidden to report done with failing output, re-read modified files before closing
• Context & Edit Safety: re-read files after 10+ messages, chunk large files, re-read before/after every edit, multi-target search before renames, one source of truth
• Judgment: flag misconceptions before executing, surface adjacent bugs proactively
• Scope: no improvements beyond what was asked, simplest approach first, plans are plans not code
• Self-Evaluation: two-perspective review, bug autopsy, failure recovery after 2 failed attempts
• Housekeeping: checkpoint offers before risky changes, flag unwieldy files
• Comment policy: default no comments, only when WHY is non-obvious

Test plan

• Docs only — no code changes