I built an LFS system when it came out years ago and it was a good learning exercise. That's the cool thing here. I wouldn't consider using the system from day to day though.
I agree with your sentiment though. A recent attempt to get sstp-client and NetworkManager-sstp working on CentOS 6.3 resulted in a long stompy walk when I couldn't get it to build without breaking the distribution due to dependency hell. I just chucked Ubuntu on the box in the end and used the vendor provided deb packages. Was much easier :-)
In that context we're really just solving a management problem with engineering! Machines are pretty great at following procedure and ensuring consistency after all.
Go has essentially solved these problems, but only for Go code and in the process threw out (for now, at least) the notion of binary linkable libraries that your code can access, which may or may not be a bad thing depending upon how you feel about dynamically linkable libraries.
I don't think the problem is practically fixable on existing languages whose basic toolset wasn't built to deal with these sort of things. And there is a cultural element in addition to the technical one -- even if you managed to create such a build toolset for languages like C/C++, some (large) subset of C/C++ programmers are going to want to hold on to their hard-won mastery of all things configure/autconf/make and stick with them, even if those tools are all completely horrible (which, IMO, they are).
Nowadays I avoid building anything from source if there is a binary I can download and everytime I have to build from source I cringe in apprehension as to whether it will work or not. Too often, if something fails, I just give up.
OK, I’m not really this aiming this criticism at the LFS project. I’m really just griping about shitty this aspect of software development in general. People should be working on solutions to these problems. Or maybe this is a problem with me. I’d be interested other’s opinions.