fix(backends): make streamed replies and session state consistent

- Validate stage prefixes before forwarding model tokens. Invalid prefixes
  may use the existing repair budget; once output starts, validation failure
  aborts the turn without appending a regenerated reply. Bound the token
  queue and close inner streams when consumption stops.
- Replace delete-time lock eviction with reference-counted session locks
  covering holders and waiters. Serialize set/get with chat/stream/delete
  and release cancelled waiters without replacing a live lock.
- Store person synonyms under workflow-native driver keys and expose
  consistent alias views to single/group reads and graph prompt injection.
- Separate committed field patches from public formUpdate visibility.
  Commit callbacks keep repository fields synchronized even when updates
  are hidden or the consumer closes after graph completion.
- Disable OpenAI SDK automatic retries. Own bounded retry attempts in the
  application, use 1/2/4 second backoff, retry streams only before text is
  emitted, enforce attempt timeouts, and close upstream/client resources.
- Translate FastGPT SDK max_retries (total attempts) from the configured
  extra-retry count, and verify the real SDK loop uses four attempts.
- Document the per-LLM budget and streaming failure semantics without
  claiming the whole multi-call HTTP turn has the same timeout.

Validation: 267 offline tests passed, including actual SDK requests through
MockTransport, stream repair/abort, timeout/close, lock races, aliases and
formUpdate visibility; existing four dependency deprecation warnings.

src/core/fastgpt_client.py

@@ -16,11 +16,11 @@ _chat_backend: ChatBackend | None = None
 _langgraph_graph = None
 _app_service: AccidentAppService | None = None
 
-# 后端超时/重试 SLA（两后端对齐；契约测试防止再漂移）。
-# FastGPT 侧：客户端每次请求总超时 + 重试次数；
-# LangGraph 侧：LLM 每次调用总超时 + 重试次数（退避为指数式，
-# 与 FastGPT 固定 1s 不同——总尝试次数与单次超时可比，见
-# test/backends/test_sla_alignment.py）。
+# 两后端缺省 60 秒、最多额外重试 3 次，1/2/4 秒退避。
+# FastGPT SDK 的 max_retries 实为总尝试数，注入时转换为 1 + 重试次数。
+# LLM 关闭 SDK 自动重试；流式发出文本后不再重试。
+# LLM 预算按单次模型调用计算，FastGPT timeout 为 HTTP SDK 超时，
+# 均不承诺多次模型调用的整个聊天回合具有相同总期限。
 FASTGPT_TIMEOUT_SECONDS = 60.0
 FASTGPT_MAX_RETRIES = 3
 
@@ -30,6 +30,7 @@ async def lifespan(app: FastAPI):
     """Manage chat backend resource lifecycle."""
     global _chat_backend, _fastgpt_client
     Config.validate()
+    llm = None
 
     if Config.AGENT_BACKEND == "langgraph":
         # 作业 2：LangGraph fake 模式不要求 FastGPT 密钥。
@@ -38,7 +39,8 @@ async def lifespan(app: FastAPI):
 
         global _langgraph_graph, _app_service
         _langgraph_graph = build_graph()
-        _chat_backend = LangGraphBackend(_langgraph_graph, llm=build_llm_service())
+        llm = build_llm_service()
+        _chat_backend = LangGraphBackend(_langgraph_graph, llm=llm)
         # 作业 6：应用服务 + 内存 repository（作业 7 换 PostgreSQL）。
         _app_service = AccidentAppService(
             _chat_backend, InMemorySessionRepository()
@@ -48,7 +50,7 @@ async def lifespan(app: FastAPI):
             api_key=Config.FASTGPT_API_KEY,
             base_url=Config.FASTGPT_BASE_URL,
             timeout=FASTGPT_TIMEOUT_SECONDS,
-            max_retries=FASTGPT_MAX_RETRIES,
+            max_retries=FASTGPT_MAX_RETRIES + 1,
             retry_delay=1.0,
             enable_logging=Config.DEBUG,
         )
@@ -61,6 +63,8 @@ async def lifespan(app: FastAPI):
         _chat_backend = None
         _langgraph_graph = None
         _app_service = None
+        if llm is not None and hasattr(llm, "aclose"):
+            await llm.aclose()
         if _fastgpt_client:
             await _fastgpt_client.__aexit__(None, None, None)
         _fastgpt_client = None