Add JA definitions for 187 synsets missing them

[fcbond]

Summary

• 70 pronoun/demonstrative synsets (wnja-77xxxxx): grammatical glosses for 1st/2nd/3rd person pronouns (私・彼・彼女…), demonstratives (これ/それ/あれ, この/その/あの, こんな/そんな/あんな, こんなに/そんなに/あんなに), interrogatives (何/誰/どこ/いつ/なぜ), universals (みんな/全部), reflexives (自分), and meta-categories (所有代名詞/疑問代名詞)
• 117 Asia/Japan-specific concept synsets (wnja-80xxxxx): prose definitions translated from English — food terms (居酒屋, おにぎり, お盆, タピオカティー, ホーカーセンター, チリクラブ…), cultural concepts (参勤交代, 初夢, 干支, 親孝行, 少子化…), language/tech terms (人工知能, 漢字, 翻訳/通訳, DVD…), and more; named entities excluded
• tweak_wnja.py: apply_corrections() now supports adding new definitions when old_value is empty, in addition to replacing existing ones
• audit/llm.py: force empty <think> block for Qwen3 models to prevent verbose plain-text reasoning consuming the token budget (fixes Stage 1 slowdown)

Review notes

All definitions are in corrections/definitions.tsv with source=audit/missing-defs and date=2026-05-30. Please check especially:

• Pronoun glosses for naturalness (currently grammatical descriptions rather than prose definitions)
• Southeast Asian food terms for accuracy
• Any wnja-80xxxxx entries that are actually named entities and should be left without definitions

Test plan

• uv run python build_wnja.py && uv run python tweak_wnja.py — verify "196 corrections applied", no WARNINGs on new entries
• Spot-check a few synsets in the output XML
• uv run pytest tests/

🤖 Generated with Claude Code