[matt-edmondson]

Fixes #197

What changed

Fuzzy filtering compares in Unicode normal form C because it goes through ktsu.FuzzySearch, but glob and regex filtering compared raw strings. A name stored decomposed (NFD), the way macOS writes file names, disappeared in Glob and Regex mode and reappeared in Fuzzy mode.

• A new ToNormalizationFormC helper returns a string that is already NFC unchanged, so ASCII and NFC text cost no extra allocation. A string that can't be normalized, such as one with a lone surrogate, is matched as it is instead of throwing.
• Glob: the filter token is normalized in ResolveGlob before the cache key is taken, so the NFC and NFD spellings of one filter share a cache entry. The text token is normalized in IsGlobMatch. Normalization happens before ExpandGlobClasses, so Caf[éè]* works with either spelling inside the class. It also covers the public AnyTokenMatchesGlobFilter / AllTokensMatchGlobFilter entry points.
• Regex: the text and the pattern are normalized at the top of DoesMatchRegex.

Tests

• NFC filter against NFD text, and the reverse, for glob with *, ? and [...], under both case sensitivities and the default match option.
• The same for regex under ByWordAny and ByWholeString, case-sensitive and case-insensitive.
• Filter over mixed NFC/NFD items returns the same items for Glob, Regex and Fuzzy.
• NFC and ASCII input come back as the same instance, and a lone surrogate is still matched.

With normalization stubbed out, 13 of the new cases fail. With the fix, every new test passes. dotnet test gives 282/283. The one failure, ReadmeTests.EveryMethodInTheReadmeApiReferenceExistsWithItsDocumentedDefaults, also fails on main (5f420d9) without this change. Branch fix/183-readme-api-test-compat-overloads appears to cover it.

---

Generated by Claude Code

[matt-edmondson]

CI on 0143ded failed ci / .NET / Test on all three OSes. The failing test was ReadmeTests.EveryMethodInTheReadmeApiReferenceExistsWithItsDocumentedDefaults. It isn't caused by this PR: the scheduled CI run on main (5f420d9) fails the same way, and the test passes or fails identically with or without this change.

A fix already exists on open branch fix/183-readme-api-test-compat-overloads (9328e00), so I've cherry-picked it here as 0aa598d rather than leave this PR red until that one merges. With it, dotnet test passes 283/283 locally.

---

Generated by Claude Code

[matt-edmondson]

The SonarCloud quality gate (Maintainability C on new code) isn't caused by this PR. SonarCloud last analysed main on 2026-11-04, at 837b460. Every scheduled CI run on main since then has failed on the README test, so that analysis never refreshed. This PR is therefore compared against a five-day-old baseline. The issues counted as "new code" include the four S3427 BLOCKERs from the TextFilter.BinaryCompatibility.cs overloads (#166) and the S3776 in ExpandGlobClasses, neither of which this diff touches. main's own analysis already lists them. The only findings on lines this PR adds are three INFO-level MSTEST0068 suggestions in the new tests. Fixing those would not get the gate to A.

What clears it is a green main that SonarCloud can re-baseline against. That is #186, which this PR already carries as a cherry-pick. Once #186 merges, re-running this PR's CI should give a fresh comparison. #200 is in the same state.

---

Generated by Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
12 New issues
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud