Timsort is very cool, and it would definitely be fun to implement in numpy. It is also a lot more work that merging two sorted arrays! I think +1 if someone else does it, but although I would love to be part of it, I am not sure I will be able to find time to look into it seriously in the next couple of months.

From a setops point of view, merging two sorted arrays makes it very straightforward to compute, together with (or instead of) the result of the merge, index arrays that let you calculate things like `in1d` faster. Although perhaps an `argtimsort` could provide the same functionality, not sure. I will probably wrap up what I have, put a lace on it, and submit it as a PR. Even if it is not destined to be merged, it may serve as a warning to others.

I have also been thinking lately that one of the problems with all these unique-related computations may be a case of having a hammer and seeing everything as nails. Perhaps the algorithm that needs to be ported from Python is not the sorting one, but the hash table...