I think you need both. You need randomised tests that will get you to cornercases. You also need fixed tests, i.e., every random test that ever failed, should be made into a fixed test. Then you are certain to avoid regressions, while still exploring the complete spectrum on every test run.

You import sortBy, yet you do not use it, instead opting for the more expensive reverse after the sort.

I think this will walk the first k elements twice. Once to take i elements off the list and once to compute last.