fix: comprehensive CJK support for pattern detection

[matsukaze-minden]

Summary

Extends #19's scope to provide comprehensive CJK (Chinese, Japanese, Korean) support for pattern detection — covering both false positive prevention and correction pattern detection.

• Support full-width ？ and CJK question particles (嗎吗呢か까) in false positive filter
• Add CJK-aware short message threshold (2 chars for CJK vs 4 for ASCII)
• Add non-correction English phrase filter (No problem, don't worry, never mind, etc.) to prevent false positives in mixed CJK-English text
• Add 13 CJK correction patterns: Japanese (8), Chinese (3), Korean (2)
• 22 new tests covering all changes

Problem

The existing pattern detection is English-centric, causing two classes of issues for CJK users:

1. False positives (English patterns triggering on non-corrections)

Input	Matched pattern	Actual intent
No problem, 次に進もう	^no[,. ]+	Agreement, not correction
don't worry、大丈夫	^don't\b	Reassurance, not correction
Never mind、別の方法でやろう	^never\b	Dismissal, not correction
何ですか？	(not matched by \?$)	Question with full-width ？

2. False negatives (CJK corrections not detected)

Input	Expected detection	Actual
いや、そっちじゃなくて	Correction ("no, not that")	Not detected
違う、useStateじゃなくて	Correction ("wrong, not useState")	Not detected
それ間違ってる	Correction ("that's wrong")	Not detected
不是，应该用另一个方法	Correction ("no, use another method")	Not detected

Changes

scripts/lib/reflect_utils.py

False positive prevention:

• FALSE_POSITIVE_PATTERNS: \?$ → [?\uff1f]$ + CJK question particles
• NON_CORRECTION_PHRASES: New list of 8 English phrases that look like correction openers but aren't (No problem, don't worry, never mind, etc.)
• detect_patterns(): CJK-aware short message threshold (≤2 chars for CJK, ≤4 for ASCII)

CJK correction detection:

• CJK_CORRECTION_PATTERNS: 13 patterns across 3 languages
  • Japanese (8): いや、違う、そうじゃなくて、間違ってる、じゃなくて〜にして、やめて、そうじゃない、って言った
  • Chinese (3): 不是、错了/錯了、不要X要Y
  • Korean (2): 아니、틀렸
• Integrated into detect_patterns() flow between false positive check and English correction check

tests/test_reflect_utils.py

• New TestCJKPatternDetection class with 22 tests:
  • Short message rejection (ASCII and CJK-aware thresholds)
  • Full-width question mark and CJK question particles (ja/zh/ko)
  • Non-correction English phrases in mixed text
  • All 13 CJK correction patterns
  • English pattern regression tests

Relationship to #19

This PR is a superset of #19. All three changes from #19 are included:

• ✅ Full-width ？ support
• ✅ CJK question particles
• ✅ Short message rejection

Plus additional improvements:

• ✅ CJK-aware short message threshold (CJK chars carry more meaning per character)
• ✅ Non-correction English phrase filter
• ✅ 13 CJK correction patterns (Japanese, Chinese, Korean)
• ✅ 22 comprehensive tests (vs 3 in fix: add CJK false positive filtering for question detection #19)

Test plan

• All 94 tests pass (72 existing + 22 new)
• No behavior change for existing English detection patterns
• Verified with real-world mixed Japanese-English prompts
• CJK-aware threshold correctly handles single-character CJK vs multi-char ASCII

🤖 Generated with Claude Code

[shohu]

Closing: will resubmit from correct account.

[matsukaze-minden]

Closing: resubmitting from correct account (shohu).

[shohu]

Could you please delete this PR? It was submitted from the wrong account by mistake. The final version is #24. Sorry for the clutter.