Test the 20% error reduction with leakage checks

Test the 20% error reduction with leakage checks

Evaluate the candidate model once on the untouched test families and require at least a 20% mean absolute error reduction against the task 2 baseline. Check that no test family, formula, or duplicate entry appears in training. Deliverable: A scorer reporting the relative reduction, a per-family breakdown, and a leakage report that fails on any shared formula.