fix(llm): 终审 I-1/I-2+M-1/2/3/5——传输异常纳入重试链出口 LLMError(502 契约)/sleep 改 asyncio.sleep(事件循环不冻结)/save sentence 上限/坏 YAML fail-soft/删死脚手架 [vps] [no-doc]

sanguo_api/hypothesis_card.py

@@ -46,12 +46,12 @@ _SYSTEM_TEMPLATE = """你是A股量化投研假设卡设计专家。用户会给
 
 
 def load_domains() -> list[str]:
-    """读 data_manifest.yaml domains 键清单;文件缺失 fail-soft 返回 []."""
+    """读 data_manifest.yaml domains 键清单;文件缺失/坏 YAML fail-soft 返回 []."""
     path = os.environ.get("SANGUO_DATA_MANIFEST", "config/data_manifest.yaml")
     try:
         with open(path, encoding="utf-8") as f:
             manifest = yaml.safe_load(f) or {}
-    except OSError:
+    except (OSError, yaml.YAMLError):
         return []
     return sorted((manifest.get("domains") or {}).keys())
 

sanguo_api/llm/client.py

@@ -11,8 +11,7 @@ from __future__ import annotations
 import asyncio
 import json
 import logging
-import time
-from typing import Callable
+from typing import Awaitable, Callable
 
 import httpx
 
@@ -29,10 +28,10 @@ _STRICT_NOTE = "Strictly output valid JSON. No extra text."
 class LLMClient:
     def __init__(self, config: LLMConfig,
                  transport: httpx.AsyncBaseTransport | None = None,
-                 sleep: Callable[[float], None] = time.sleep) -> None:
+                 sleep: Callable[[float], Awaitable[None]] | None = None) -> None:
         self._cfg = config
         self._transport = transport
-        self._sleep = sleep
+        self._sleep = sleep or asyncio.sleep
 
     async def chat_json(self, messages: list[dict], *, temperature: float = 0.2,
                         max_tokens: int = 2000) -> dict:
@@ -48,7 +47,14 @@ class LLMClient:
         attempts = 0
         while True:
             attempts += 1
-            status, text = await self._post(payload)
+            try:
+                status, text = await self._post(payload)
+            except httpx.TransportError as e:
+                # 传输级故障(连接拒绝/超时/断连)视作一次失败 attempt,同款退避
+                if attempts >= _MAX_ATTEMPTS:
+                    raise LLMError(f"LLM 连接失败: {e}") from e
+                await self._sleep(_BACKOFF_SECONDS[min(attempts - 1, 1)])
+                continue
             if status == 200:
                 content = self._content(text)
                 usage = json.loads(text).get("usage") or {}
@@ -76,14 +82,17 @@ class LLMClient:
             if status == 429 or status >= 500:
                 if attempts >= _MAX_ATTEMPTS:
                     raise LLMError(f"LLM 重试耗尽(HTTP {status}): {text[:200]}")
-                self._sleep(_BACKOFF_SECONDS[min(attempts - 1, 1)])
+                await self._sleep(_BACKOFF_SECONDS[min(attempts - 1, 1)])
                 continue
             raise LLMError(f"LLM 未预期状态 HTTP {status}: {text[:200]}")
 
     async def _strict_retry(self, messages: list[dict], payload: dict) -> dict:
         payload = dict(payload, messages=messages + [
             {"role": "user", "content": _STRICT_NOTE}])
-        status, text = await self._post(payload)
+        try:
+            status, text = await self._post(payload)
+        except httpx.TransportError as e:
+            raise LLMError(f"LLM strict 重试连接失败: {e}") from e
         if status != 200:
             raise LLMError(f"LLM strict 重试失败(HTTP {status}): {text[:200]}")
         try:

sanguo_api/routes_pipeline.py

@@ -844,6 +844,9 @@ async def hypotheses_draft(body: dict) -> dict:
 
 @router.post("/pipeline/hypotheses", status_code=201)
 def hypotheses_save(body: dict) -> dict:
+    sentence = str(body.get("sentence") or "")
+    if len(sentence) > _DRAFT_SENTENCE_MAX:
+        raise HTTPException(400, f"sentence 超长(>{_DRAFT_SENTENCE_MAX} 字)")
     snake = {
         "title": body.get("title"),
         "logic": body.get("logic"),