Build a cost-derived confidence router

Keep three things separate: thresholds() is pure arithmetic from costs to band edges; decide() applies hard rules, then the confidence floor, then the per-action lines on $p_{\max}$; and the evaluation (risk–coverage and realised cost) runs over a labelled set. Unknown or never-automated intents have no lines at all, so a new intent added to the Choice escalates by default. The offline run uses a seeded simulated model that is calibrated by construction, so you can see what good probabilities do to the table before trying a real backend.