Cover CJK and four-byte characters in the filename tests

[jakejackson1]

Nothing under tests/ carried a Han ideograph, kana or Hangul syllable, and the only four-byte character was an emoji in FilenameTest::provideInvalidUtf8ToSanitize — the display sanitizer's provider — so no four-byte character was ever driven through setName().

Adds 26 tests. No source change.

providerSetNameSanitizing — runs in every mbstring configuration

Row	Case
82	聊天(PC)-201711201615.xlsx, CJK mixed with the ASCII punctuation the rewriter does touch
83	Katakana ファイル名.txt
84	Hangul 파일이름.txt
85	CJK Extension B 𠮟𪚲.txt, four-byte
86	photo😀.png, four-byte
87	文字(v2).文 — an extension is ASCII alphanumeric or discarded whole
88, 89	A three- and a four-byte character ending exactly on the 250 byte budget, kept

Row 82 is the shape a client name off a CJK desktop arrives in: three-byte characters either side of parentheses the rewriter turns into -. Nothing tested that combination.

providerSetNameSanitizingRequiringMbstring — @group mbstring

Every offset at which a wider character can straddle the truncation cut: pads 248 and 249 for a three-byte character, 247 through 249 for a four-byte one. The provider had one two-byte case at pad 249.

This is the half with a branch behind it. forceValidUtf8() matches an incomplete tail with a separate alternative per width — [\xE0-\xEF][\x80-\xBF]? and [\xF0-\xF4][\x80-\xBF]{0,2} — and the one existing case exercised neither. The rows land on different code by configuration:

• ext-mbstring loaded — mb_strcut() refuses to split the character, so they test its boundary handling.
• mbstring polyfill job — that job runs without --exclude-group, and the polyfill ships no mb_strcut(), so finalize() falls to substr() and the repair strips the partial sequence.

Rows 88 and 89 are in the plain provider on purpose: a character ending on a byte boundary survives a byte cut unchanged, so the expected value holds with no repair and every configuration can run them. Verified against the substr() path rather than assumed.

Checks

• Suite 628 → 654 tests, 1270 → 1322 assertions, still exactly one skip, so cross-file-system's --fail-on-skipped guard is intact.
• lint, phpstan (level 9) and check-syntax clean.
• Non-vacuity checked by mutation: mb_strcut → mb_substr fails 16 tests, adding \x80-\xFF to the unsafe-character class fails 62.

No CHANGELOG.md entry — test-only, and tests/ is export-ignored.

Run locally with ext-mbstring loaded. The polyfill and no-mbstring configurations are CI-only.

🤖 Generated with Claude Code