A Pipeline That Does Not Leak

The single most common mistake in applied ML, and it always makes the numbers look better.

Write fit_and_score(X, y) that:

1. Builds a pipeline which imputes missing numeric values with the median and standardises them
2. Fits logistic regression on it
3. Returns the mean accuracy from 5-fold cross-validation

The trap: if you impute and scale before splitting, every fold's training set has already seen the median and standard deviation of its own validation set. The reported accuracy goes up and means nothing: the model is being told something it will never know at prediction time.

The preprocessing must be fitted inside each fold, on that fold's training data only.