We use Unreal Build Accelerator which injects a custom dlib into clang to intercept the compilation process and distribute it to worker machines.

I’m curious why this needs to be code injection and not, like, a shell script?

Because...it's official behaviour that is fully supported by clang? If you want to add a hook on compilation start, it's literally the documented way - you include your own dlib with necessary overrides and then you can call your own methods at each compilation step. Not even sure how you'd do it with a shell script? You need to have knowledge of all the compilation and linking units, which....you have from within Clang.