Problem: experiments show raw counts only — no significance (audit: src/handlers/experiments.ts).
Scope: compute conversion rate + confidence/p-value (or Bayesian) per variant; flag winners.
Done when: results indicate whether a variant is significantly better.