fix: 客户端响应结构损坏归入 LLMResponseError；3xx 不再误判成功

src/genesis/inference/client.py

@@ -1,11 +1,17 @@
 from __future__ import annotations
 
+import json
 import time
 from typing import Protocol, Sequence
 
 import httpx
 
-from .exceptions import LLMNetworkError, LLMNotConfiguredError, LLMTimeoutError
+from .exceptions import (
+    LLMNetworkError,
+    LLMNotConfiguredError,
+    LLMResponseError,
+    LLMTimeoutError,
+)
 from .types import ChatMessage, TokenUsage
 
 
@@ -81,9 +87,16 @@ class HttpLLMClient:
                 continue
             if resp.status_code >= 400:
                 raise LLMNetworkError(f"LLM HTTP {resp.status_code}: {resp.text[:200]}")
+            if not (200 <= resp.status_code < 300):
+                # 3xx 重定向不自动跟随，不得误判为成功
+                raise LLMNetworkError(f"LLM HTTP {resp.status_code}: {resp.text[:200]}")
 
-            data = resp.json()
-            content = data["choices"][0]["message"]["content"]
+            try:
+                data = resp.json()
+                content = data["choices"][0]["message"]["content"]
+            except (json.JSONDecodeError, KeyError, IndexError, TypeError) as exc:
+                # 2xx 但响应结构损坏（非 JSON / 缺字段）→ 结构化错误，不重试
+                raise LLMResponseError(f"LLM 响应结构损坏: {exc}") from exc
             usage_raw = data.get("usage", {})
             usage = TokenUsage(
                 input_tokens=usage_raw.get("prompt_tokens", 0),