feat(eval): Playground GLM 评分落库并可恢复

在 conversations 表增加 playground_conversation_judge_json，流式/非流式对话评审结束后写入最近一次快照（整体分、逐轮分、对比文案、错误与基线文件名等）。新增只读 GET 供前端按会话拉取；评测台 Playground 切换会话时自动恢复，并提示基线是否和当时一致。

api/app/features/evaluation/router.py

@@ -46,6 +46,7 @@ from app.features.evaluation.schemas import (
     ManualJudgeConversationStreamBody,
     ManualJudgeMemoirBody,
     ManualJudgeMemoirOut,
+    PlaygroundConversationJudgeOut,
     MemoirPhase1ReadyOut,
     MemoirSectionBaselineOut,
     RegressionSetCreate,
@@ -225,6 +226,26 @@ async def get_session_transcript(
     )
 
 
+@router.get(
+    "/sessions/{conversation_id}/playground-conversation-judge",
+    response_model=PlaygroundConversationJudgeOut,
+)
+async def get_playground_conversation_judge(
+    conversation_id: str,
+    _auth: InternalEvalAuth,
+    db: Annotated[AsyncSession, Depends(get_async_db)],
+):
+    catalog = SessionCatalogService(db)
+    tr = await catalog.get_transcript(conversation_id)
+    if not tr:
+        raise HTTPException(status_code=404, detail="conversation not found")
+    judge = await catalog.get_playground_conversation_judge_json(conversation_id)
+    return PlaygroundConversationJudgeOut(
+        conversation_id=conversation_id,
+        judge=judge,
+    )
+
+
 @router.get(
     "/sessions/{conversation_id}/memoir-phase1-ready",
     response_model=MemoirPhase1ReadyOut,