Regression in rustdoc of re-export involving namespace overlap

[dtolnay]

// src/lib.rs

mod thing {
    pub use core::sync::atomic::AtomicU8; // anything from a dependency, type namespace

    #[allow(non_upper_case_globals)]
    pub const AtomicU8: () = (); // value namespace
}

pub use crate::thing::AtomicU8;

cargo +nightly-2022-12-02 doc	cargo +nightly-2022-12-03 doc

I believe the one on the left is what I would consider the correct rendering, and the one on the right is a regression.

The regression is in nightly-2022-12-03, specifically #104963 according to bisect, which makes sense. Mentioning @petrochenkov @cjgillot in case an obvious cause stands out to either of you.

[petrochenkov]

in case an obvious cause stands out to either of you.

In short, rustdoc middle-end is not written well, nothing is obvious in it, and changes that are easy and predictable in rustc may lead to entirely unpredictable consequences in rustdoc.
In this specific case I investigated the affected rustdoc code just enough to make the test suite pass.

[dtolnay]

Mentioning @GuillaumeGomez who has been recently touching many of the same files modified in #104963 (for example in #103886 and #105183).