You're absolutely right. Both this one and Aldous-Broder both have a worst-case where the algorithm never terminates. As for whether the algorithm is "unacceptable", that depends on the application. For games? Yeah, this is probably far from your best option for generating mazes. But for cases where you absolutely must have a uniform spanning tree, your options are limited. Wilson's is much better than Aldous-Broder, but still not perfect. Robin Houston has described a variant that combines both Aldous-Broder and Wilson (doing AB until about 30% of the field is filled, and then switching to Wilson's) which empirically improves the odds quite a bit, but as long as you're doing a blind random walk, you're pretty much never guaranteed to finish.
By the way, thanks for this series of articles... I've found it most diverting. I guess I'm the sort of person who enjoys "recreational maze generation" :p
