Oops, my bad. Thanks for catching that ikegami. The info was from benchmark notes a few years old, so I don't really know what happened.

My best guess is that the benchmarks shown for 10,000 key hashes were actually done on empty hashes. I ran those again and updated the post (also checked the other results). From this, the while-each and foreach loops appear to gain the speed advantage as the hashes get larger.