Description mentions caching the results, but it is not tested. Either this should be a requirement and hence tested, or should be removed from the description. If you require caching, how should void functions be dealt with? The description could benefit from clearer guidelines.

The test only tests up to 3 arguments, which can lead to easy non-variadic solutions. This definitely shouldn't be allowed.

Ruby 3.0 should be enabled, read this to learn how to do it

Please organize the structure of test fixture as following (although it has been mentioned in the attached link, I'm repeated here again)

describe "<This message should describe the categories of test groups inside this block>" do
  it "<A short message describing this test group>" do
    expect(...).to eq(...) #Assertions
    #Or Test.assert_equals(user_response, reference_response)
  end
end

Call call no more than once!

Users are going to implement call, so it shouldn't be called by them. Did you mean that call won't be called more than once?

We create a function

If it's a requirement, it should be tested. For instance, my solution fails at

auto f = [] { };
make_lazy(f);

because make_lazy isn't a function.

Probably a duplicate: https://www.codewars.com/kata/modular-multiplicative-inverse
Or is there an algorithm that can handle 150000000 in time, but not 2^64?

Unless you've done the other kata, there's no way of having any clue on what you're meant to be doing. I'd suggest explaining what the coder needs to do to pass the kata, in your description.

This kata is dead simple because you can generate all permutations.
It will be a lot more interesting in base 36 or in base 64 where you can't iterate on all permutations before timing out.
Besides 10000 tests is too much and really not useful (my current solution wiping them in 900ms): if you have tricky numbers to test, use a random set of fixed tests.

Isn't 10,000 tests a bit excessive? 1,000 would be more than enough, and that way it wouldn't choke the server.. I can't see this kata being passable on slow CW days.

python:

That might be a problem... Though, it makes the current rank estimation correct.