fix: 诊断提速与稳定性修复（reasoning 2048→512、输出精简、上限16000→8000、loading提示、观测日志+配置错误可读）

test_demo.py

@@ -113,10 +113,18 @@ def test_reasoning_config_does_not_set_effort_and_max_tokens():
     retry = reasoning_config(True)
     assert not ("effort" in first and "max_tokens" in first)
     assert not ("effort" in retry and "max_tokens" in retry)
-    assert first == {"max_tokens": 2048}
+    assert first == {"max_tokens": 512}
     assert retry == {"effort": "none", "exclude": True}
 
 
+def test_bad_proxy_url_raises_readable_llm_error():
+    # 代理 URL 带变量名前缀（如 .env 解析出错）→ httpx 构造期 ValueError；
+    # 必须转可读 LLMError，不能裸 500（2026-08-25 实测事故，见 llm.py 注释）
+    c = LlmClient(api_key="test-key", proxy="PRODREAM_BACKEND_OPENROUTER_PROXY_URL=http://x:1@h:2")
+    with pytest.raises(LLMError, match="配置错误"):
+        c._complete("system", "user")
+
+
 def test_message_text_empty_mentions_finish_reason():
     with pytest.raises(LLMError, match="finish_reason=length"):
         message_text(