To me, legacy code has cost. Great cost.
You can't just leave old code alone and the features remain. Any time you go to restructure some code, or support some other mode, or do some other kind of action, it ends up affecting this old code and you have to go and make changes to it as well. Then that old code breaks on a new version of the compiler, so you have more fixes to make. And it increases compilation times, binary sizes, and lines of code. It's another vector for malicious attacks. It's more to keep mentally in your head all at once.
I suppose it's one thing if you have a steward willing to take over, say, the a.out branch of the binary loader. But it's quite another in my case where I'm the sole developer of my software, and no one shows any interest in maintaining the legacy UI, or versions of the emulation core that focus more on performance at the expense of compatibility, so I have to drop that stuff. (Yet it is a bit disappointing when people are willing to fork the entire codebase instead to achieve the same goals, but I digress.)
Again, I agree with what you're saying, and I'd say Linux or Mozilla Firefox or Qt certainly has the manpower for this sort of thing. But as a solo developer, I can see why people would sacrifice this backward-compatibility from time to time.
I'm not sure on your point about getting a feel for interfaces. Certainly, it seems that every year or two, I feel like I've made great strides and am finally on the cusp of stable interfaces. Yet invariably, a year or two later, I look back and think the same thing again. I've come to the conclusion that we never stop evolving and that thinking we've reached perfection is always self-delusion. But, maybe I'll get there in the future. I certainly hope so.