Trade Ranking Versus Probability Quality

Model A has slightly higher ROC AUC, but its probabilities are visibly overconfident; Model B has slightly lower ROC AUC but much better calibration. If you need a probability to size positions according to expected payoff, which model is usually safer?