TLDR: scikit-learn's 'LogisticRegression' function violates the principle of least astonishment by automagically using L2 regularization.