That argument baffled me as well. Using the 'dictionary size' defined by the distinct set of characters as an input for the entropy calculation seems absurd, unless the attacker has a way to know the dictionary, for instance because they know a login can be made using a numeric keypad alone.

Also, with that definition, a password 'abcde' would have much higher "entropy" than 'aabab', meaning that this measure penalizes actual randomness, where repetition and char-reuse are very likely.