TBH, I am off two minds here, as the following concerns conflict:

• as a library author, I want to pull as much complexity as posible into the lib's code, so that, from the user's perspective, things "just work".
• at the same time, I very much don't want to "go against the grain of the language", and implement clever work-arounds.

I guess, on balance, yeah, we should do that.

But before, I am curious, can we optimize this a bit? In particular, the current impl,

    pub struct Lazy<T, F = fn() -> T> {
        cell: OnceCell<T>,
        init: Cell<Option<F>>,
    }

isn't really what it should be, as we have two flags here one in OnceCell, and one in Option. And we don't share the space between T and F.

So I think ideally we want something like

struct Lazy<T, F> {
    state: enum { Init, Uninit, Poison },
    data: union {
        f: F,
        t: T,
        poison: (),
    }
}

can we make that work pedantically correct wrt variance?

I guess, as another consideration here, give that std::sync::LazyLock is almost stable, I'll probably change what this crate is about, and move it from "stable foundation" to "fancy extras", as the foundational role would be subsumed by std. From this perspective, doing something more creative here also makes sense.

can we make that work pedantically correct wrt variance?

Yes, by slightly amending it 🙂:

use ManuallyDrop as MD; // <- necessary red tape

struct Lazy<T, F> {
    state: SomeMutWrapper<enum { Init, Uninit, Poison }>, // Uninit=can read `f`, Init=can read 't', Poison=cannot read either
    data: union {
        f: MD<F>,
        t: MD<UnsafeCell<T>>, // 👈 `UnsafeCell` only here
    }
}

• e.g., the impl Drop:

  match self.state.get_mut() {
      Init => unsafe { MD::drop(&mut self.data.t) },
      Uninit => unsafe { MD::drop(&mut self.data.f) },
      Poison => {},
  }

• takeing the f would then be a matter of setting the state to Poison, then <*const _>::read()ing the .f field like this very PR does, and writing that value to .t, and then setting the state to Init.

This has the right variance w.r.t. both F and T and would indeed allow squashing the is_some flag of .f in my PR within state. The only drawback is that it requires reïmplementing the flag-handling for Lazy, rather than being able to rely on the OnceCell<T> already existing abstraction.

• The advantage, however, for people using Lazy with zero-sized impl Fn types (e.g., within function bodies or using TAITs), it that it would mean Lazy would be getting smaller for them as well!

  In other words, this last implementation you've suggested would be both pedantically correct w.r.t. variance, and efficient!

I'm thus closing this draft PR, but may be submitting a PR implementing this for unsync::Lazy (the "easy" to write implementation of OnceCell). From there, I'll let sync-savy people follow-up on this and handle the sync flavor thereof