And this is why the kernel is "everywhere" while "year of the desktop" never happens. Because Linux userspace is all to happy to break itself again and again and again...
Isn't this more of a social problem than a technical one? It seems to me like the main issue is a convention that a system only have a single version of a library installed, the anointed "system" version of that library. And often, binary releases of software are built against a particular cocktail of library versions from a particular distro at the time of release, each of which depends on its own cocktail of libraries, etc... In reality it's not complicated to get virtually any binary running on any system (in the worst case, just put the distro it expects into a chroot) - and there's no reason why applications couldn't just bundle all the libraries they need, right down to glibc - but it's just Not Done.
That was an interesting read; he wasn't even very confrontational there. The point was also solid - they had an internal performance optimization change which was unproven, yet broke things in the real world. That, and the lack of guidelines or tests for what kind of usage results in the behaviour, it's a de-facto "feature" of the interface that they should not break.
I do wonder what the correct workflow is for preventing this - release binaries with assertions in place for development, so that if any of your input constraints are violated that aren't guaranteed to fail fast in the release version, then it's enforced in this debug version?