I must be missing something here. Using e.g. Mercurial, I would start my refactoring and commit often. If a test fails at the end and I can't see why then just update to the version before the refactoring (assuming I hadn't ever ran tests in the mean time), run the test, if it passes update to the next revision and so on. I would expect this to take seconds or minutes to isolate. What kind of revisions system is he using that this even comes up?

You're right that frequent commits greatly simplifies the particular problem of undetected errors in big refactorings. However, people being people, the sunk cost fallacy is still an issue in development.