According to condition (1), this is the desired behavior. I added a sentence to (2) to make this clearer:

"NOTE: When the array is both not square and the wrong size, (1) applies, so "Array not square" should be returned."

Great, now it works :) I made some final adjustments to the description and naming, but now it is approved and published.

Hm, not really. Try to add a print(matches(*i)) in the end of the random tests and you'll see that all tests still evaluate to False. Like this:

for i in random_ips:
    print(matches(*i))
    test.assert_equals(matches(*i), reference_solution(*i))

Read my last post carefully and see if you can make 50% of the random tests evaluate to True.

The way I do it is to randomize test data until I get variables that makes the match() function return true, the next return value must be false, then true and so on. If you randomize a wildcard mask that looks like this 00000000.00001111.11111111.11111111 the likelyhood if getting variables that return true is much greater.

Have a look at my tests Test(match, random) { } under the C Kata Reference.

I'm a bit unsure if it should be translated into Python, but I have approved it. It sure is slightly harder in C though.

I'm not sure about C but this is not a 6kyu kata in Python.

I think your first line . . .

    arr2 = arr

. . . only creates a reference to arr, then all changes on arr2 will be reflected on arr too.
arr2 = arr[:] would be more efficient to your purpose.
; )