It seems like you're assuming that software cannot change without a rewrite. Just look at the Linux kernel and compare what it was at the time of 1.0 vs 3.0. The two versions have nothing in common. Yet, nobody ever started a rewrite of the kernel and the only regressions that occurred were either unintentional (i.e. bugs happen no matter what) or were about hardware stopped using 10 years earlier. It's been an evolution and that's how you keep your users happy. It took years for gnome2 to re-implement many of the features that gnome1 had. Some It's not clear to me whether these features were always meant to be re-implemented or were only re-implemented because of angry users, but the bottom line is that for gnome2, "change" resulted in several years of being stuck with something worse. Not to mention that gnome1 had stopped improving long before gnome2 was out. So far, gnome3 appears to be the same. gnome2 development slowed down long before the gnome3 release and developers are still re-implementing features that were "lost" in the rewrite (again don't know how much is from "it took time" vs "we changed our mind and re-implemented the feature").
The bottom line is that this development method (and KDE is just as guilty as gnome) leads to long periods where the software is much worse than it should be. So by "learn from gnome2" I meant "learn to avoid these long periods of regression". I now realize that "it's a feature" and that even if gnome3 ever becomes as usable for me as gnome2 was before the "rewrite", it would only be temporary because by the time it works for me, all developers will have moved to gnome4, which will break the features I use.
I especially don't see how it needs to be that way. You can implement a gnome-shell like feature on top of gnome2 and make it optional (or even by default, I don't care). You can make the wm evolve without breaking everything. And more importantly, you don't have to make all these changes at once and you don't have to ditch the old behaviour. This means 1) people have time to get used to the changes when they like it 2) there's a chance to react when you're going in the wrong direction, and 3) you avoid bad regressions.