That's a really cool solution. One potential problem is that it doesn't find the loop as soon as possible (e.g. if it's going around the perimeter of a square, the cycle is detectable after 4N steps, but the algorithm will take N^2 steps to realize that)
That's nice (and can be easily fixed to handle negative coordinates), but extremely slow (imagine a simple square cycle with side length N. Her solution takes N^2 steps.
Perfection! A beautiful hack around not knowing the size of the board, and even faster detecting cycles to boot.
It's an extension of your idea of counting the total number of visited positions. What I like most about it is that it can work with streamed data.