The key is the program which generates this program lol

This is so tough for me, and very enlightening. I think it's worth a 1kyu, at least no less than any other 1kyu kata lol.

If I understand the join of stream correctly, it takes the diagnol infinite matrix generated by the stream of streams, like the one here:

instance Monad Stream where
  return = repeat
  xs >>= f = join (fmap f xs)
    where
      join :: Stream (Stream a) -> Stream a
      join ~(Cons xs xss) = Cons (head xs) (join (map tail xss))

so how could I encounter (2, 3) in the product of two S.gen (fun n -> 1 + n) (L.return 0)s?

8 completions and no ratings, which is amazing.

Actually there does not seem any description of base91 on the internet. I don't even know where to get started, apart from reading the source code of some implementations.

If it doesn't check the invalid cases, then why pointing them out?

You can wait for the author to approve it :-)

I seem to have done a translation in Scala, but it's too ugly in that it requires many type hints.

The Java version does not have a description.

Didn't it went wrong when checking for "Integer division"?

Approved :-)

But I have to say that if the answers to my questions are both "omitted", things would be more interesting and realistic. (Wouldn't be much harder though)

Looks like my guess is correct. If the algorithm works and no error message is output, then no timeouts.