Preserve letters outside the BMP in every case conversion

[matt-edmondson]

Fixes #70.

The bug

All three regexes tested Unicode categories per UTF-16 code unit. A letter outside the Basic
Multilingual Plane is a surrogate pair, and each half is UnicodeCategory.Surrogate rather than a
letter — so [^\p{L}0-9] matched both halves, ReplaceNonAlphaNumericWithSpace's predecessor
turned the letter into two spaces, and CollapseSpaces removed them. The letter was gone.

Measured at main (5a20678), .NET SDK 10.0.401, Linux:

input	before	after
"set 𐐀abc".ToPascalCase()	SetAbc	Set𐐀abc
"𐐀abc".ToSnakeCase()	abc	𐐨abc
"abc𝒳def".ToMacroCase()	ABC_DEF	ABC_𝒳DEF
"HELLO 𐐨".IsAllCaps()	true	false

Two further faults shared the root cause, and both are fixed here:

• IsAllCaps skipped astral letters. It stripped non-letters with [^\p{L}] — which removed
  both surrogate halves — then tested the remainder, so an astral lowercase letter was never seen.
  ToTitleCase uses that result to decide whether to lowercase before title-casing.
• The splitter inserted a spurious word boundary. (?<=[\p{L}])(?=[^\p{L}]) read the high
  surrogate as a non-letter, so abc𝒳def split into two tokens.

Changes

All three regexes are replaced by walks over code points:

• ReplaceNonAlphaNumericWithSpace — keeps letters and ASCII digits, replaces everything else
  with a space.
• SplitOnCaseChange / IsWordBoundary — the regex's three alternatives, expressed as explicit
  conditions over code points. The comments name the case each one covers.
• IsAllCaps — walks code points directly rather than stripping and re-testing.

ToTitleCase gains Ensure.NotNull(input), which preserves the ArgumentNullException that
Regex.Replace used to throw and satisfies CA1062 now that the regex call is gone. The
#if NET7_0_OR_GREATER split disappears with the regexes, so all six target frameworks now run
identical code.

Astral characters that are not letters — emoji, symbols — remain separators, exactly as before.

Verification

• Behaviour for BMP input is unchanged. The new implementation was compared against the
  original regex implementation over 12,191 inputs (all three-part combinations of 23 atoms
  covering ASCII case, digits, separators, punctuation, and accented/Greek/eszett letters, plus 24
  realistic identifiers) across all six public methods — no differences. The same harness was
  then re-run with an astral input added and correctly reported the four expected differences, so
  it is not vacuously passing.
• New tests fail without the fix. Reverted the source change and re-ran: 4 of the 5 new tests
  fail, 25 pass. The fifth pins that non-letter astral characters are still dropped, and passes
  either way by design.
• Full suite: 29/29 pass. The 24 pre-existing tests are unchanged.
• Build: clean across net10.0, net9.0, net8.0, netstandard2.0, netstandard2.1,
  0 warnings, 0 errors.

Not in this PR

"MAX_SIZE".ToPascalCase() is MaxSize but "set MAX_SIZE".ToPascalCase() is SetMAXSIZE — the
same token converting two ways depending on what precedes it. That falls out of ToTitleCase
testing IsAllCaps on the whole string to decide whether words are acronyms, which is a
documented heuristic rather than an oversight. Changing it is a design decision about acronym
policy, so it is left alone and noted on #70's sibling discussion in ktsu-dev/Coder#71.

🤖 Generated with Claude Code

https://claude.ai/code/session_014RUsrYfSAuSQ7VWws7vzsr

---

Generated by Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
97.5% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud