There is a fallacy which says "if I understand the steps which got me here, than I understand where I am." Or "If I understand each change that has been made to a system, then I understand the system." This is simply not true. A myopic understanding of each change does not equal an understanding of the system. Human memory is simply not good enough to remember all of the previous bits of code, and comprehend whether a change won't happen to interact in a bad way.

With lots of testing, yes, you can incrementally build a working system that you don't understand. But by myopically staring at diffs all day, you'll only get farther from the truth of how your system works.

The major refactors that I do, are because I've come to the realization that my data model was not correct. A realization, that often happens multiple times during a project. My utility functions are still useful, but I have to rewire everything. There is no meaningful way to do that incrementally. Why would you even try to do that incrementally? It's like, if you have a salamander and you want to turn it into a horse incrementally. Of course, most of the RNA is the same. But testing and code reviewing the steps between salamander and horse seems to me like a waste of time.