One-hot encoder with fitted categories

The problem, from first principles

A model cannot multiply a weight by the word "red", so categorical features must become numbers. One-hot encoding gives each category its own column: a row holds a 1 in its category's column and 0 everywhere else. The category list must be learned from training data, and an unseen category at prediction time should be reported, not silently encoded as all zeros.

Implement two functions.

• one_hot_fit(values): returns the sorted list of distinct values, which fixes the column order.
• one_hot_transform(values, categories): returns a float matrix of shape (len(values), len(categories)). Column j corresponds to categories[j]. A value not in categories raises ValueError.

Constraints

• Empty input returns an array of shape (0, len(categories)).
• Each row of a valid output sums to exactly 1.
• The categories passed to one_hot_transform define the column order, even if they are not sorted.

Hints

Open one at a time. Each gives away a little more than the last.