feat(prompts): implement TOML-backed prompt loading and refactor message generation

python/ebook_search/protected_phrases/judge_ngrams.py

@@ -14,6 +14,7 @@ from sqlalchemy import select
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from python.ebook_search.llm_interface import request_chat_completion
+from python.ebook_search.prompts import load_prompt
 from python.ebook_search.protected_phrases.extraction import (
     candidate_source_names,
     get_sample_contexts,
@@ -369,20 +370,7 @@ def build_judge_messages(candidate: PhraseCandidate) -> list[dict[str, str]]:
         "chapter_count": candidate.chapter_count,
         "contexts": candidate.sample_contexts,
     }
-    return [
-        {
-            "role": "system",
-            "content": (
-                "Judge whether a candidate phrase from a book should be protected for RAG retrieval. "
-                "Do not extract new phrases. Reject common grammar fragments, ordinary nonspecific phrases, "
-                "unstable fragments, and phrases kept only because they are frequent. Keep people, places, "
-                "organizations, factions, events, technologies, fictional conditions, magic systems, formal titles, "
-                "named concepts, and recurring world-specific terms. Return only a JSON object with keys: keep, "
-                "canonical, category, aliases, confidence, importance, allow_nested, suppress_children, reason."
-            ),
-        },
-        {"role": "user", "content": json.dumps(payload, ensure_ascii=True)},
-    ]
+    return load_prompt("phrase_judge").messages(candidate_json=json.dumps(payload, ensure_ascii=True))
 
 
 def parse_llm_judgment(content: str, config: EbookSearchConfig) -> LLMJudgment: