It's because it has to work in pattern contexts as well, which are not expressions.

They tested the same strings on that implementation

The code they were looking at was used for writing the table, but they were testing the one that read it (which is instead correct).

though judging by the recent comments someone’s found something.

Yeah that's me :)The translation using an associated const also works when the const block uses generic parameters. For example:

fn require_zst<T>() {
    const { assert!(std::mem::size_of::<T>() == 0) }
}

This can be written as:

fn require_zst<T>() {
    struct Foo<T>(PhantomData<T>);
    impl<T> Foo<T> {
        const FOO: () = assert!(std::mem::size_of::<T>() == 0);
    }
    Foo::<T>::FOO
}

However it cannot be written as:

fn require_zst<T>() {
    const FOO: () = assert!(std::mem::size_of::<T>() == 0);
    FOO
}

Because const FOO: () is an item, thus it is only lexically scoped (i.e. visible) inside require_zst, but does not inherit its generics (thus it cannot use T).