fix(yoga): Chinese combination text from a Chinese source, not 未判定 (#91 fix2 F3, BUG-1418)

Removing the English-to-Chinese back-translation left every BVR semantic
note and every yogas_doshas raj/dhana "H*_Lord(...) in H*" sentence as
未判定 in the Chinese yoga table and pages.

- yoga_engine: semantic conditions return BiText(Chinese, old note). The
  Chinese comes from references/yoga_semantic_notes_zh.json, one sentence per
  note, so dedup keeps the same equality classes. English-only custom
  templates get the same lookup. combination_en is unchanged.
- yogas_doshas: raj/dhana rows add combination_zh; combination is unchanged.
- chinese_combination prefers a clean combination_zh.

Which yogas fire, their strength and category are unchanged (fictional
charts 0-19, same rows as staging; only set-iteration order differs).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HwMNZEzdzmrbPhuRNmZe3C

scripts/report_render/appendix_language.py

@@ -647,6 +647,10 @@ def chinese_yoga_label(item: dict) -> str:
 
 def chinese_combination(item: dict) -> str:
     """Use the Chinese combination sentence. An English sentence is not translated word by word."""
+    authored = str(item.get("combination_zh") or "").strip()
+    if authored and re.search(r"[一-鿿]", authored) and not _has_prose_glue(authored):
+        # Written in Chinese by the detector itself (yogas_doshas raj / dhana rows).
+        return authored
     combination = str(item.get("combination") or "").strip()
     han = bool(re.search(r"[\u4e00-\u9fff]", combination))
     function = bool(_HAN_FUNCTION_RE.search(combination)) if combination else False