Nondeterminism in ignore::WalkBuilder parallel multi-root walk

[jelle-openai]

Please tick this box to confirm you have reviewed the above.

• I have a different issue.

What version of ripgrep are you using?

14.1.1

How did you install ripgrep?

Cargo

What operating system are you using ripgrep on?

macOS 26.5

Describe your bug.

ignore::WalkBuilder appears to produce nondeterministic results for a parallel multi-root walk when one root has a scoped ignore rule that should not apply to another root.

What are the steps to reproduce the behavior?

Minimal layout:

.git/
.gitignore
noxfile.py
src/scikit_build_core/build/metadata.py
tests/test_metadata.py

.gitignore:

tests/**/build/

The source build file should always be included: the ignore rule only matches build/ directories below tests/.

Minimal Rust repro with ignore = "0.4.25":

use std::path::Path;
use std::sync::atomic::{AtomicBool, Ordering};

fn source_build_file_was_seen(root: &Path) -> bool {
    let source_build_file = root.join("src/scikit_build_core/build/metadata.py");
    let seen = AtomicBool::new(false);

    let mut builder = ignore::WalkBuilder::new(root.join("src"));
    builder.current_dir(root);
    builder.standard_filters(true);
    builder.add(root.join("tests"));
    builder.add(root.join("noxfile.py"));
    builder.threads(12);

    builder.build_parallel().run(|| {
        Box::new(|entry| {
            if let Ok(entry) = entry {
                if entry.path() == source_build_file {
                    seen.store(true, Ordering::Relaxed);
                }
            }

            ignore::WalkState::Continue
        })
    });

    seen.load(Ordering::Relaxed)
}

fn main() {
    let root = Path::new(env!("CARGO_MANIFEST_DIR"));
    let mut seen = 0;
    let mut missing = 0;

    for _ in 0..100 {
        if source_build_file_was_seen(root) {
            seen += 1;
        } else {
            missing += 1;
        }
    }

    println!("seen={seen} missing={missing}");
}

Observed output on my machine:

seen=14 missing=86

The same issue reproduces via ripgrep:

rg --threads 12 --files src tests noxfile.py

Across 100 identical runs, src/scikit_build_core/build/metadata.py was present 59 times and missing 41 times.

Control cases were stable:

rg --threads 1 --files src tests noxfile.py
# source file present 100/100

rg --threads 12 --files src
# source file present 100/100

So this looks specific to the parallel multi-root walk. The expected behavior is that tests/**/build/ never excludes src/scikit_build_core/build/metadata.py, regardless of walk scheduling.

This is causing nondeterminism in ty too: it makes it so some files in the scikit-build-core project are sometimes checked and sometimes ignored by ty. The repro case in this issue is a simplified version of the shape in scikit_build_core.

What is the actual behavior?

Inconsistent output from rg --threads 12 --files src tests noxfile.py.

What is the expected behavior?

Rules should be applied consistently