Big old code bases, have a lot of what I call as 'tribal knowledge' associated with them. People go out, and do experiments only to burn their hands. Over years, the old maintainers generally accumulate a lot of ideas and 'tribal knowledge' to know why 'pie in the sky' ideas won't work.

Nevertheless, people still try. A lot of them fail, a few succeed.

The vim code base is around 300K lines of C code. I'm sure its nothing like Java where 90% of the code is get/set methods, high level delegating classes and exception handling code. 300K lines of code which does the actual work is simply too much code. And one should exercise great caution while dealing with it. Regression is just one thing, I'm sure there are a lot of places in the code which can silently break things you can't detect only to pop up in some strange case in the wild.