Wire the ECE metric into a calibration report (or remove it)

`eval/metrics.py` defines expected calibration error (`ece`) but nothing references it except a unit test — it feeds no committed report.

**Scope (good first issue):** either (a) surface `ece` in a small committed calibration report alongside the other eval outputs, or (b) remove `ece` and its test. Pick one and make the codebase consistent.