Current description is indecipherable and there are too much missing information:

• Input list have at most depth 3
• 1st level is always a list; 2nd level can be a int or a list of ints
• Find all lists resulting from exploding 2nd level of each list; for every list of ints in 2nd level, they should be exploded as set product; then list all by order of appearance in the result
• Result should be a list of lists; sub-lists should be right-padded with 0 to the longest sub-list in the result
• Sub-lists should have minimal length 1

• Test framework has to be used correctly: Sample tests still have top level assertions, and imports for solution and codewars_test are missing.
• Description does not specify whether inputs can contain anything other than ints and None. Is it necessary to skip strings when searching for a nearest int?
• Suggestion: input arrays are quite short, so test inputs could be easily presented to a user on failure to make debugging easier.

Test cases would have a better shape and would be more readable and maintainable if they were factorized. I suggest to create a function dotest that handles all test cases. Then all the fixed tests could be grouped in a list [(input, expected)], without repeating test.assert_equals x times.

This kata definitely consists of two tasks:

1. Figuring out, what bad means.
2. Solving performance issues.

First part is a nice puzzle. As for the second one, it's pretty easy to come around with a fixed array of bad integers. This array can either be hardcoded or calculated outside of the function. Both these ways feel like cheating to me.

So it's kinda weird: you have some performance problems to deal with, but they can be solved with the same computational complexity (linear) as the dumb algorithm, just changing the order a little bit.

IHMO, you should either remove the performance part from this kata (so that the stupid solutions like sum(1 if is_bad(i) else 0 for i in range(a, b+1)) work), or require a solution of logarithmic complexity.

Or maybe you can split this kata in two: the first one will be very simple allowing the most obvious solution, but the second one should pose the performance problem to its full.

It should be stated whether all options should be taken or only one (and which one exactly), e.g. what's the result of these:

a = [1, A]    b = [0, X]    t = 1
                  [1, Y]
                  [2, Z]


a = [2, A]    b = [0, X]    t = 2
                  [3, Y]