feat(evals): judge rubrics + statistical power analysis

reasoning-quality + behavioral-pairwise judge rubrics (consumed by the
runners) and power.js (Wilson CI, sign/McNemar, required-N — shows the
108-problem set is underpowered; ~194 needed for a 10pp aggregate).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>