feat: stream searches and add reproducible benchmarks

[ParthJadhav]

Summary

• make searches genuinely streaming with bounded memory, prompt cancellation, exact concurrent limits, and error-aware traversal
• fix literal filename matching, strict searches without extensions, directory leakage, case-insensitive extensions, and repeated metadata reads
• add lossless path results, multiple extensions, result-type selection, minimum depth, symlink, filesystem-boundary, thread-count, and early file-size controls
• improve similarity sorting and update maintained dependencies while removing the direct regex dependency
• add controlled and cross-tool benchmarks, expanded regression coverage, stronger CI, an MSRV, changelog, benchmark report, and product roadmap

Performance

On the same warm 100,000-file fixture, with equal 10,000-path outputs and 15 measured runs:

Revision	Mean	Range
This PR	57.1 ms	52.3–68.2 ms
260286c	78.1 ms	71.8–86.2 ms

That is a 1.37x throughput improvement.

The reproducible 50,000-file tool comparison verifies identical sorted sets of 5,000 paths before timing:

Tool	Mean	Range
rust_search	30.7 ms	29.9–33.2 ms
ripgrep --files	31.0 ms	30.1–32.2 ms
fd	35.0 ms	33.2–37.7 ms
find	95.9 ms	88.1–99.0 ms

See BENCHMARKS.md for the environment, methodology, caveats, internal timings, and reproduction commands.

Compatibility

This corrects two behaviors that callers may have observed despite the documented API:

• default searches now consistently return files and links instead of sometimes including directories
• search_input is now literal text instead of implicitly interpreting regex punctuation

Search construction also returns before traversal completes. These changes are documented in CHANGELOG.md; maintainers should consider whether the next publication warrants a major version.

Verification

• 49 unit and integration tests
• 24 doctests
• macOS symlink and traversal tests
• Linux-only non-UTF-8 path regression test in CI
• cargo fmt --all -- --check
• cargo clippy --all-targets -- -D warnings
• rustdoc with warnings denied
• Rust 1.88 MSRV check
• packaged-crate compilation
• exact cross-tool result-set parity before benchmarks

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 59bb0a46d7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Cancel dropped searches before finding a match

When a query has no matches in a large tree, dropping the iterator only disconnects the result receiver, but this branch never observes that disconnection and continues walking every entry. Since channel failure is checked only while sending a match or traversal error, rare- or no-match searches can leave detached walker threads consuming filesystem I/O long after callers stop using them, contrary to the documented prompt cancellation behavior.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Escape extensions before installing the glob prefilter

On filesystems that permit glob metacharacters in extensions, inputs such as .ext("?") or .ext("*") are interpreted as glob patterns rather than literal suffixes. TypesBuilder::add accepts these patterns, after which extension_prefiltered causes the exact extension comparison to be skipped, so ? matches every one-character extension and * matches arbitrary extensions; the prefilter must escape user input or retain the literal matcher.

Useful? React with 👍 / 👎.