The test is very weak; there are no actual random tests except one test with randomized parameters, so it's easy to specifically pattern match against it (considering that the operations are very easy to notice, this is not very hard).

There is a issue with random test when the string starts with h.
According to the description, it should be all caps when the input string starts with h.

But the following input string failed when the output is all caps

hcUR72Dm6xdSaNjoz Y6xFXqx 4 1Qu M TpC

The test expected this: OMG HCUR72DM6XDZANJOZ Y6xFXqx 4 1Qu M TpC but as I said, according to the description it should be like this: OMG HCUR72DM6XDZANJOZ Y6XFXQX 4 1QU M TPC

Javascript should have a wider test coverage to invalide this solution with an incorrect word length count.

My solution successfully passed all predefined tests, but sometimes fails on some random tests (and sometimes not).

I suspect it's because there are no predefined test cases that verifies that replacement should be case-sensistive, for example, 'S' is always replaced with 'Z' and not 'z'. Random tests somtimes have that cases.

I like this kata, but the description of the problem should be clearer e.g. (the priority of each rule)

The (python) "random test" selects 10 cases from a pool of 5 predefined examples -- so, there is nothing random there, literally...

Suggestions: Have test cases for rules in the order they're described (disregarding full validity, i.e. 'partial credit').

The every-other-word caps, omg/lol (albeit not quite as much), and the question-marks/exclamation-points are where I expect to debug. It would be nice to have everything else clearly validated in the meantime.