Claude Code arena for 100-agent code duels and blind judging

Build me a /arena tournament for Claude Code: 1. Check ~/.claude/agents and .claude/agents for agents that fit architect, adversary, inversion and judge roles > Create only the missing ones as subagent files > Workers on Sonnet 5.5, the lead evaluator and the judge on Opus 5.5 2. Create .claude/commands/arena.md: spawn N solvers in waves of 10, each in its own git worktree under .arena/runs/, with no shared context 3. Pair them head-to-head: each attacks the other's solution at three points (before code runs, when input scales, before done), then one revision round, then a blind judge 4. Write rubric.md: 100 points. Correctness 30, completeness 25, robustness 20, specificity 15, clarity 10. A fatal flaw is an automatic loss 5. Add one rule to CLAUDE.md: changes to critical architecture go through the arena before merge Show me every diff first. Don't apply anything until I confirm.