bug#22990: Grafts leads to inefficient substitute info retrieval

[David Craven]

> On second thought, the whole idea of applying grafts on the final result
> (instead of applying grafts at each step like we do now) doesn’t fly.
> It works well for things like a profile or the system derivation, but
> breaks for less trivial things.

> For example, if you’re building a VM image or a binary tarball, you
> really need to graft packages early on; trying to graft the VM image or
> binary tarball wouldn’t have the desired effect.

Isn't a system derivation or a profile derivation an intermediate step
to these derivations? Can't there be a flag or something called
#:already-grafted? #t?