feat: LLM/Embedding 重试次数纳入 system.yaml 统一配置

- system.yaml: llm.max_retries(死配置) 改为 llm.max_attempts；embedding 段新增 max_attempts
- llm/client.py: LLMConfig 新增 max_attempts（默认 3 兜底），load_llm_config 读取配置
- llm/extractor.py: translate_and_extract(_async) max_attempts=None 时取 config.max_attempts
- embedding/client.py: load_embedding_config 读取 embedding.max_attempts
- scheduler/reporter.py: _call_llm_simple max_retries=None 时取 max_attempts-1
- 新增 2 个配置读取测试；已同步 pi5 验证

configs/system.yaml

@@ -43,10 +43,10 @@ dedup:
 # ── LLM 翻译+事件抽取 ────────────────────────────────
 llm:
   provider: "deepseek"
-  deepseek_model: "deepseek-chat"
+  deepseek_model: "deepseek-v4-flash"
   qwen_model: "qwen-plus"
   timeout_sec: 60
-  max_retries: 3
+  max_attempts: 3          # 单篇总尝试次数（含首次），失败后指数退避重试；日报 AI 摘要同用此值
   max_tokens: 8192
   temperature: 0.1
   concurrency: 3
@@ -57,6 +57,7 @@ embedding:
   dashscope_model: "text-embedding-v3"
   dimension: 1024
   batch_size: 10
+  max_attempts: 3          # 单批总尝试次数（含首次），失败后指数退避重试
   timeout_sec: 30
 
 # ── Qdrant ──────────────────────────────────────────