Another minor change worth calling out - I added a note to the docs that when a symlink cycle is detected, realpath() could return any member of the cycle, but does not guarantee which one.

For our test cases it's generally stable enough, but if you change the starting point (e.g. from absolute to relative path) then it can change the result. It seemed easier all round just to not guarantee anything about which member of the cycle will be returned, rather than determining the exact rules, reproducing them across all platforms, and documenting them.