Their main benchmark test is compiling every publicly released crate on crates.io. This is also their main regression test.

If you manage to overfit against that, it's still probably an amazing general purpose solution.

Yes that's likely. But the idea is that even if that's the case, it is still better than no PGO.

Edit: I'd like to add that if the 10-20% mentioned is measured on the benchmark that was used to do the pgo, then that figure might indeed not be representative of the real gain.