I think the issue is that the zero-cost abstraction terminology is used in two ways often. First, to mean a feature you can use and costs nothing because the compiler will change it into a more common form for you (such as functional programming approaches where it will take multiple filters and map passes and turn it into a single loop), and second, to mean that features added that do add overhead, only do so when you use them (as in, there's no runtime to manage M:N threading unless you include a crate that does so, because developers not using that feature shouldn't have to pay for it).

I think the terminology here is looser than it should be, and these are not really the same thing, even though I've seen zero-cost abstractions used to describe them both. It causes confusion sometimes, as I believe it is doing here.