Unfortunately, adding a couple more of fixed tests is not a complete solution, because fixed tests can be easily hardcoded. While a fixed test will point you to a flaw in your solution, it will not force you to fix it, but can be satisfied just by hardcoding it. The solution would be to add fixed tests and fix random tests so they are guaranteed to generate inputs of all kinds. And there is already many issues about this (I counted 7 issues related to insufficient tests).

You can propose a fork with more fixed tests, but this is only a partial solution. If you created a fork with good random generators for all types of expected results, it would be great.

The random tests in Python don't cover much and their value isn't random afaik. There is an open issue about that saying there are no random tests, closing this one. Adding a couple more of fixed tests isn't the solution. I guess a much wider range should be checked the question is how.

This is specified in the description:

NOTE 2: The 0x0 (empty matrix) is represented as en empty array inside an array [[]].

I don't think so. the group by conditions are different - there is no reusable logic here