Improve validation in Section 5.1

[reetahan]

1. Calibration-vs-prediction passage
2. held-out moments: Re-run calibration using only a subset of statistics as targets (e.g., top-3 match rate, unmatched rate). Report fit against statistics \emph{not} used as targets (top-5, top-10, gender breakdowns, RBO structure)
3. naive baselines for Figure 2