Build a calibration toolkit and test it on two models

The metrics are pure functions over two lists, probabilities and 0/1 outcomes, so they work the same on simulated data, on jevlocal output and later on real Jev answers. Check them against the lesson's worked example and TypeSafe's documented confidence examples before trusting any number they print. The decision that matters most is the simulator: a calibrated model reports the true probability $q$, and an overconfident one reports $\sigma(2\,\text{logit}(q))$, which pushes every probability away from 0.5 without changing which side of 0.5 it is on.