feat(api): DeepSeek V4 Flash 默认、HTTP 错讯与多供应商分层

- 主链路默认 deepseek-v4-flash，DEEPSEEK_THINKING_ENABLED 对齐旧非思考 chat
- 评测台评审装配迁入 adapters/llm（deepseek_eval_judge、zhipu_eval_judge）与 eval_judge_spec
- 拆分 llm_http_openai_chat_errors 与 llm_errors（DeepSeek/智谱品牌与文档链），llm_call 支持 http_error_vendor
- EvalJudgeService 按 spec.provider 传入 allm_json_call；评测台前端文案改为 V4 Flash
- 更新 .env 示例与 staging/production 的 DEEPSEEK_MODEL；补充 openai/供应商错讯测试

Made-with: Cursor

api/app/features/evaluation/judge_service.py

@@ -7,6 +7,7 @@ from dataclasses import dataclass
 from typing import Any, Generic, TypeVar
 
 from app.core.config import settings
+from app.core.eval_judge_spec import EvalJudgeProvider
 from app.core.llm_call import LLMCallError, allm_json_call
 from app.core.logging import get_logger
 from app.features.evaluation.judge_schemas import (
@@ -284,8 +285,10 @@ class EvalJudgeService:
         judge_llm: Any | None,
         *,
         context_window_tokens: int | None = None,
+        http_error_vendor: EvalJudgeProvider = "deepseek",
     ) -> None:
         self._llm = judge_llm
+        self._http_error_vendor: EvalJudgeProvider = http_error_vendor
         self._ctx_tokens = int(
             context_window_tokens or settings.eval_judge_context_window_tokens
         )
@@ -329,6 +332,7 @@ class EvalJudgeService:
                 TurnJudgeOutput,
                 max_tokens=_TURN_MAX,
                 agent="EvalJudgeService.judge_turn",
+                http_error_vendor=self._http_error_vendor,
             )
         except LLMCallError as e:
             logger.warning("turn judge failed: {}", e)
@@ -354,6 +358,7 @@ class EvalJudgeService:
                 ConversationJudgeOutput,
                 max_tokens=_CONV_JUDGE_JSON_MAX,
                 agent="EvalJudgeService.judge_conversation",
+                http_error_vendor=self._http_error_vendor,
             )
             return JudgeCallResult(output=out)
         except LLMCallError as e:
@@ -496,6 +501,7 @@ class EvalJudgeService:
                     512, int(settings.eval_judge_memoir_completion_max_tokens)
                 ),
                 agent="EvalJudgeService.judge_memoir",
+                http_error_vendor=self._http_error_vendor,
             )
             return JudgeCallResult(output=out)
         except LLMCallError as e: