You compile a version of the dependency which intentionally behaves different (e.g. introduces random network errors, random file parsing errors, etc.) and "inject" it into a otherwise functional setup to check if all other parts including error reporting and similar work.

There are always other ways to archive this but using (abusing?) the dynamic linking is often the simpler way to set it up.

> and what aspect of this is useful for simulators and solvers?

Duno, but some programs allow plugin in different implementations of the same performance critical functionality so that you can distribute a binary version and then only re-compile that part with platform specific optimizations. If that is what the author is referring to I would argue today there are better ways to do it. But it's still working out either way and can be much easier to setup/maintain in some cases. (And probably falls into the "shared libraries as extension system" category Linus excluded from his commentary.)