feat(mcp): 基座换行分帧与 isError 映射；新增 dwg_review 审查 Agent

基座（cad_mcp_frame）:
- 新增 RecvOneFrame: 以 '\n' 为消息边界分帧，去掉单次 recv(2048) 限制（上限 8MB）；
  用 select 短探测兼容不发 '\n' 的旧客户端；clientSocket 设 SO_RCVTIMEO 防卸载卡死
- JSON-RPC id 支持 string/number；无 id 的通知不回包
- isError 映射: 协议级错误保留顶层 error，工具执行错误转 result.isError=true
- 响应补 '\n' 作为消息边界

Agent（新增 dwg_review 包）:
- mcp_client: 唯一 TCP/JSON-RPC 封装（单飞锁/分帧/重试/兼容 error 与 isError）
- cad_tools: 从 tools/list 动态建表，含 15 个只读白名单
- phases: planning -> evidence -> review（受控只读取证子循环）-> report
- annotator: 批注回写 AI_REVIEW 层，支持 dry-run 与撤销
- config: 配置全部走环境变量；Windows 下回退读注册表以读到新设的系统变量

其它: 新增 verify_mcp_frame.py 验证脚本；.gitignore 补 Python/AutoCAD 锁文件规则；
ai.memory 记录架构决策与实测结论

dwg_review/llm_client.py

@@ -0,0 +1,101 @@
+"""LLM 封装（文本 + 视觉），统一超时 / 重试 / 截图预处理。
+
+默认走 DeepSeek（视觉模型 deepseek-flash）。密钥从环境变量读取，禁止硬编码。
+"""
+
+import base64
+import io
+import time
+
+from openai import OpenAI
+
+from . import config
+
+try:  # 可选依赖：仅用于截图下采样
+    from PIL import Image
+
+    HAS_PIL = True
+except ImportError:  # pragma: no cover
+    HAS_PIL = False
+
+
+class LlmError(Exception):
+    pass
+
+
+def downsample_image_b64(b64, mime="image/png", max_edge=None, quality=None):
+    """把 base64 截图下采样并转 JPEG，降低视觉 token 成本。
+
+    缺少 Pillow 或参数为 0 时原样返回 (b64, mime)。
+    """
+    max_edge = config.IMAGE_MAX_EDGE if max_edge is None else max_edge
+    quality = config.IMAGE_JPEG_QUALITY if quality is None else quality
+    if not HAS_PIL or max_edge <= 0:
+        return b64, mime
+    try:
+        img = Image.open(io.BytesIO(base64.b64decode(b64)))
+        if img.mode not in ("RGB", "L"):
+            img = img.convert("RGB")
+        w, h = img.size
+        scale = max_edge / float(max(w, h))
+        if scale < 1.0:
+            img = img.resize(
+                (max(1, int(w * scale)), max(1, int(h * scale))), Image.LANCZOS
+            )
+        buf = io.BytesIO()
+        img.save(buf, format="JPEG", quality=quality)
+        return base64.b64encode(buf.getvalue()).decode("ascii"), "image/jpeg"
+    except Exception:
+        return b64, mime
+
+
+class LlmClient:
+    """统一的 LLM 调用入口。"""
+
+    def __init__(self, api_key=None, base_url=None, model=None):
+        key = config.API_KEY if api_key is None else api_key
+        if not key:
+            raise LlmError("未配置 DEEPSEEK_API_KEY 环境变量")
+        self.model = model or config.MODEL_NAME
+        self._client = OpenAI(api_key=key, base_url=base_url or config.BASE_URL)
+
+    def chat(self, messages, *, tools=None, timeout=None, retries=None):
+        """发起一次对话补全，返回原始 response。"""
+        retries = config.LLM_RETRIES if retries is None else retries
+        last = None
+        for attempt in range(retries + 1):
+            try:
+                kwargs = {
+                    "model": self.model,
+                    "messages": messages,
+                    "timeout": timeout or config.LLM_TIMEOUT,
+                }
+                if tools:
+                    kwargs["tools"] = tools
+                return self._client.chat.completions.create(**kwargs)
+            except Exception as exc:  # noqa: BLE001 - 统一包装为 LlmError
+                last = exc
+                if attempt < retries:
+                    time.sleep(0.5 * (2 ** attempt))
+                    continue
+                raise LlmError(f"LLM 调用失败: {exc}") from exc
+        raise LlmError(f"LLM 调用失败: {last}")
+
+    def chat_text(self, system, user, **kwargs):
+        """纯文本对话，返回助手文本。"""
+        resp = self.chat(
+            [
+                {"role": "system", "content": system},
+                {"role": "user", "content": user},
+            ],
+            **kwargs,
+        )
+        return resp.choices[0].message.content
+
+    @staticmethod
+    def image_content(b64, mime="image/png"):
+        """构造 OpenAI/DeepSeek 视觉消息里的图片内容块。"""
+        return {
+            "type": "image_url",
+            "image_url": {"url": f"data:{mime};base64,{b64}"},
+        }