Conformal prediction sets for triage

Calibrate on one split, evaluate on another. calibrate() returns the $\lceil (n+1)(1-\alpha) \rceil$-th smallest score (or infinity when there are too few points), and prediction_set() keeps every label with $1 - p \le \hat q$. The synthetic model is deliberately overconfident and weak on the rare bank_details_change intent, so the per-intent table shows what "marginal" coverage hides.