Random access impl of `Take::for_each`/`Take::fold` suppresses side effects of `next` call on empty iterator

[maxdexh]

I tried this code:

use std::fmt::Debug;

fn main() {
    eprintln!("does_print:");
    consume(does_print());
    eprintln!("does_not_print:");
    consume(does_not_print());
}

fn does_not_print() -> impl Iterator<Item: Debug> {
    [1, 2, 3].iter()
}
fn does_print() -> impl Iterator<Item: Debug> {
    [1, 2, 3].iter().filter(|_| true) // use any noop adapter without `TrustedRandomAccessNoCoerce` here
}

fn consume(iter: impl Iterator<Item: Debug>) {
    // the skip covers all the elements, meaning `take` is called on an empty iterator
    iter.map(|it| dbg!(it)).skip(3).take(100).for_each(|_| {});
}

I expected to see this happen:

Either both iterations print, or neither does, since the only difference is that one iterator is being adapted with filter(true), which is a noop.

Instead, this happened:

does_print:
[src/main.rs:11:19] it = 1
[src/main.rs:11:19] it = 2
[src/main.rs:11:19] it = 3
does_not_print:

@rustbot label A-iterators T-libs

[maxdexh]

@Indrapal-70 the command is @rustbot claim

[Indrapal-70]

@rustbot claim

[workingjubilee]

See #161676 (comment)

I believe that Take is subject to the same underlying reasoning problems that Zip was, though it may not have had its contract specified to work identically (or its contract is just underspecified...) in which case we might need to discuss this anew.

cc @the8472

[maxdexh]

@rustbot label A-docs

[maxdexh]

I think that both zip and take should not beat around the bush by saying stuff like "... at most one time"
Instead, they can just explicitly state that in certain (unspecified) cases, they may skip calls to empty iterators, even if that suppresses side effects