Heapsort implementation vary widely in their cache coherence. The vast majority are extremely simple but result in essentially complete incoherence. I was once able to speed a heapsort about 2x by rearranging the scan order into a cache-oblivious one (Mostly the heap building part; I'm not aware of a way to make the extraction part cache-friendy).

Heapsorts, quicksorts and bubble sorts have the ability to sort just the "top-n" and stop there (unlike mergesort), which is often useful, and a significat speed up (goes from nlog(n) to n+mlog(n) to get top m, or from n^2 to n times m for bubble sort) - I wish library routines actually provided that as part of a standard interface. Perhaps you could be the torchbearer in your publication?

Personally, I think just about every standard library except APL/J/K got sorting wrong. The primary sort operation should be "compute stable ordering permutation", with "sort this array" (basically only available operation in most libraries) being at most a shortcut when you don't care about the permutation.