Two things were missing: classes_ attribute (set on fit) and decision_function() (returns log-odds, i.e. the linear predictor). Many sklearn meta-estimators check for these before deciding how to handle a classifier. Also fixed predict() to return int instead of bool.