This improves the algorithm for good_size. The main observation is that division by 2 is cheap, so we can alternate between multiplying by 3 and dividing by 2 to reduce the search space to only numbers near n. This also adds an early return when n matches a composite number.
You can play around with the micro-benchmarks here (this is good_size2).
This also adds a variant for real transforms with 5-smooth numbers and exports both of them to python as a single function.