fix(knowledge): 首席架構師 Review 修復 C1+C2+I1+I2 (71→~88/100)

C1: IKnowledgeRepository Protocol 補齊 save_embedding + semantic_search +
    list_unembedded_entries，恢復 Interface 先行保護層

C2: embed_all_entries Service 層 raw SQL 移至 Repository.list_unembedded_entries()
    Service 改透過 Protocol 呼叫，符合 leWOOOgo 積木化原則

I1: asyncio.create_task 加入 _pending_tasks set 持有引用，防 GC 回收與
    Shutdown 時 Task 遺失；task done 後自動 discard

I2: OllamaEmbeddingService 從每次 new 改為 KnowledgeService.__init__ 注入，
    單一實例重用

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

apps/api/src/repositories/interfaces.py

@@ -272,6 +272,25 @@ class IKnowledgeRepository(Protocol):
         """view_count +1"""
         ...
 
+    async def save_embedding(self, entry_id: str, embedding: list[float]) -> bool:
+        """儲存向量 embedding (768 維, pgvector)"""
+        ...
+
+    async def semantic_search(
+        self,
+        query_embedding: list[float],
+        limit: int = 10,
+        threshold: float = 0.5,
+    ) -> list[tuple["KnowledgeEntry", float]]:
+        """語意搜尋 — cosine similarity, 回傳 (entry, score) 降序"""
+        ...
+
+    async def list_unembedded_entries(
+        self,
+    ) -> list[tuple[str, str, str]]:
+        """列出尚未產生 embedding 的條目 [(id, title, content)]"""
+        ...
+
 
 @runtime_checkable
 class IPlaybookRepository(Protocol):

apps/api/src/repositories/knowledge_repository.py

@@ -184,6 +184,17 @@ class KnowledgeDBRepository:
         )
         return result.rowcount > 0
 
+    async def list_unembedded_entries(self) -> list[tuple[str, str, str]]:
+        """列出尚未產生 embedding 的條目 [(id, title, content)]"""
+        from sqlalchemy import text as sa_text
+        result = await self.db.execute(
+            sa_text(
+                "SELECT id, title, content FROM knowledge_entries "
+                "WHERE embedding IS NULL AND status != 'ARCHIVED'"
+            )
+        )
+        return [(row.id, row.title, row.content) for row in result.fetchall()]
+
     async def save_embedding(self, entry_id: str, embedding: list[float]) -> bool:
         """儲存向量 embedding (768 維)"""
         # 直接用 raw SQL 寫入 pgvector 欄位