Searching for `Type::prefix` does not prioritize methods on `Type` over doc aliases that start with `prefix` on a different type

[zachs18]

Code

// lib.rs of a normal cargo project `cargo new --lib rustdoc_repro`
pub struct Foo;

impl Foo {
    pub fn bar_first() {}
    pub fn bar_last() {}
}

pub struct Quux;

impl Quux {
    #[doc(alias = "bar_first")]
    pub fn phooey1() {}
    #[doc(alias = "bar_etc")]
    pub fn phooey2() {}
}

Reproduction Steps

• cargo +nightly doc --open
• Search for Foo::bar

Expected Outcome

When searching Foo::bar, Foo::bar_first and Foo::bar_last should be prioritized over Quux::phooey1 and Quux::phooey2 (if those are shown at all).

Actual Output

The methods are shown in this order in the search results:

• bar_etc - see rustdoc_repro::Quux::phooey2
• rustdoc_repro::Foo::bar_last
• bar_first - see rustdoc_repro::Quux::phooey1
• rustdoc_repro::Foo::bar_first

Version

$ rustdoc +nightly --version --verbose
rustdoc 1.91.0-nightly (69b76df90 2025-08-23)
binary: rustdoc
commit-hash: 69b76df90c7ea63b5350d1865f92902a0b27c9a2
commit-date: 2025-08-23
host: x86_64-unknown-linux-gnu
release: 1.91.0-nightly
LLVM version: 21.1.0

@rustbot label A-rustdoc-search

Additional Details

Related to #140968, but in that example there is no leading type, and that issue has been since before 1.87.0.

Bisected: the behavior changed in #144476, before that count_ones didn't show up in the search results for BTreeSet::pop at all.

This is minimzed from the stdlib nightly docs, where I searched for BTreeSet::pop and got 18 different popcnt - see integer::count_ones methods before BTreeSet::pop_first and BTreeSet::pop_last (which were what I was looking for).

[lolbinarycat]

possible duplicate of #140968, although it sounds like you want aliases ranked lower even if they have a slightly lower edit distance...

we should really use a proper fuzzy search algo, currently we're just using standard edit distance, which is very much optimized for correcting typos and not for searches (importantly, it does not account for the fact people frequently deliberately search by substrings, and instead expects that people intend to type out the full identifier and only accidentally don't complete it, and ranks things accordingly)