perf(tools): 工具服务单实例并发（ServeTool 协程化）+ mcp-py 同改

垂直扩：单个 mcp 实例内每个工具调用分发到独立 goroutine/task 并发处理，
不再受 NATS 单订阅回调串行所限；配合队列组多副本即「单实例并发 × 副本数」横向扩。

- bus.ServeTool：回调立即起 goroutine 并返回（不阻塞 NATS 投递），信号量限并发
  （MCP_TOOL_CONCURRENCY，默认 16），handler panic → 回错误结果避免调用方干等超时。
- mcp-py mcp_gateway：_on_call 改 asyncio.create_task 派发 + Semaphore 限并发（同默认 16）。
- 测试 TestConcurrentToolServe：slow 工具阻塞时 fast 工具仍返回（旧串行下会超时）。

验证：单测通过；mcp-go/mcp-py live 重启工具就绪，工具往返正常（memory/history/kb_search
均正常响应，无 panic）。全模块 build+vet+test 全绿。

注：下游共享后端（Milvus/Neo4j/PG/嵌入端点）是横向扩的最终上限，届时扩这些而非 mcp 实例。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

sundynix-mcp-py/src/sundynix_mcp_py/mcp_gateway.py

@@ -47,6 +47,8 @@ class McpGateway:
         self.interpreter = CodeInterpreter() # Docker 隔离沙箱
         self._nc: nats.NATS | None = None
         self._tools: dict[str, callable] = {}
+        # 单实例工具并发上限（垂直扩）；队列组多副本再叠加水平扩。
+        self._sem = asyncio.Semaphore(int(os.getenv("MCP_TOOL_CONCURRENCY", "16")))
 
     async def register_tools(self) -> None:
         """注册工具名 → 处理协程。"""
@@ -77,25 +79,30 @@ class McpGateway:
         await asyncio.Event().wait()
 
     async def _on_call(self, msg) -> None:
-        """解析 ToolCall → 路由到工具 → Respond ToolResult。"""
-        try:
-            req = json.loads(msg.data)
-        except Exception as e:  # noqa: BLE001
-            await self._reply(msg, ok=False, error=f"bad tool call: {e}")
-            return
-        tool = req.get("tool", "")
-        task_id = req.get("task_id", "")
-        args = req.get("args") or {}
-        log.info("[mcp_py] tool=%s task=%s args=%s", tool, task_id, args)
-        fn = self._tools.get(tool)
-        if fn is None:
-            await self._reply(msg, ok=False, error=f"unknown tool: {tool}")
-            return
-        try:
-            content = await fn(args)
-            await self._reply(msg, ok=True, content=content)
-        except Exception as e:  # noqa: BLE001
-            await self._reply(msg, ok=False, error=str(e))
+        """立即派发到独立 task 并发处理，不阻塞订阅消息循环（nats-py 单订阅回调本是串行 await）。"""
+        asyncio.create_task(self._handle(msg))
+
+    async def _handle(self, msg) -> None:
+        """解析 ToolCall → 路由到工具 → Respond ToolResult（限并发）。"""
+        async with self._sem:
+            try:
+                req = json.loads(msg.data)
+            except Exception as e:  # noqa: BLE001
+                await self._reply(msg, ok=False, error=f"bad tool call: {e}")
+                return
+            tool = req.get("tool", "")
+            task_id = req.get("task_id", "")
+            args = req.get("args") or {}
+            log.info("[mcp_py] tool=%s task=%s args=%s", tool, task_id, args)
+            fn = self._tools.get(tool)
+            if fn is None:
+                await self._reply(msg, ok=False, error=f"unknown tool: {tool}")
+                return
+            try:
+                content = await fn(args)
+                await self._reply(msg, ok=True, content=content)
+            except Exception as e:  # noqa: BLE001
+                await self._reply(msg, ok=False, error=str(e))
 
     @staticmethod
     async def _reply(msg, *, ok: bool, content: str = "", error: str = "") -> None: