The kata does not make sense on making user code guess the threshold. There is no way to "guess" a threshold that is 100% correct, which is what the kata is asking us to do; what we're doing at best is coming up with flimsy heuristics that will pass the kata's tests, but will fail against many real-world examples (e.g with high ISO there will be a lot of noises, so a lot of bright dots; they might even be brighter than the object of interest), and if you make the test cases too tight, then it approaches the territory of "guessing the heuristics reference solution uses", which is equally undesirable.

Error messages are useless.

You can use something like

expected = ...
actual = crop_image(image)
message = f'{actual} should equal {expected}'
test.expect(np.array_equal(actual, expected), message)

Rust translation ready for review.

EDIT

Unpublished while I revisit the random test generator. I've found something weird.