feat(v2): 架构与算法优化二轮——推理机不变量外提、知识库匹配预编译、协作循环增量索引

- inference.py（/chat/legacy 热路径）：kb.match 循环不变量外提（原每步全量重扫+重排序，
  最坏 O(steps×rules×patterns)）；fired 查重 list→set
- knowledge.py：Rule patterns 注册侧懒缓存小写副本（原每条规则每次匹配重复 lower）；
  match() 文本只 lower 一次（原逐规则重复）；load() 的 yaml 文件名集合提到循环外
- worker.py：本地端点生成器 httpx.AsyncClient 懒建复用（原每步新建/销毁连接，
  对齐 ArchitectClient 惯用法；协作循环最多 10 次生成免重复建连）
- pipeline.py（协作循环）：plan_by_id O(1) 步定义查找；done 集合增量维护
  （原每轮重建 progress+archive 扫描）；领域只解析一次（原 _artifact_name 每步
  全领域 kb.match）；_deps_done 支持传入预填集合（保持旧签名兼容）
- v2stats.py：回合数分布改增量聚合（sum/max/分桶计数），summary() O(n)→O(1)，
  不再持有无界 list（修长时运行内存增长）
- gateway/agent.py + api.py：AgentService 运行计数 O(1) 化（原 register 全量扫描），
  状态迁移收敛到 _transition_state 单一入口（api.py cancel/异常两处绕过点一并接入，
  消除计数与状态脱节隐患）；21 项 agent 测试全绿（两轮全量 230 passed 复核）

gateway/agent.py

@@ -312,6 +312,13 @@ class AgentService:
         self.run_dir = Path(run_dir)
         self._runs: Dict[str, AgentRunInfo] = {}
         self.max_running = 5
+        self._running = 0  # O(1) 运行计数（原 register 每次全量扫描 _runs）
+
+    def _transition_state(self, info: AgentRunInfo, new_state: str) -> None:
+        """状态迁移的单一入口：离开 running 时同步递减计数。"""
+        if info.state == STATE_RUNNING and new_state != STATE_RUNNING:
+            self._running -= 1
+        info.state = new_state
 
     # ---------- 路径 ----------
     def _dir(self, request_id: str) -> Path:
@@ -327,14 +334,14 @@ class AgentService:
     def register(self, request_id: str, task: str, model: str, pool_id: str,
                  workspace: str = "", executor_model: str = "",
                  mode: str = "single") -> Optional[AgentRunInfo]:
-        running = [r for r in self._runs.values() if r.state == STATE_RUNNING]
-        if len(running) >= self.max_running:
+        if self._running >= self.max_running:
             return None
         info = AgentRunInfo(request_id=request_id, task=task, model=model,
                             pool_id=pool_id, workspace=workspace,
                             executor_model=executor_model, mode=mode,
                             started_at=time.time())
         self._runs[request_id] = info
+        self._running += 1
         self._dir(request_id).mkdir(parents=True, exist_ok=True)
         self._write_status(info)
         return info
@@ -413,7 +420,7 @@ class AgentService:
                 throttle.flush("executor")
             self._apply_result(info, result)
         except Exception as exc:  # pragma: no cover
-            info.state = STATE_FAILED
+            self._transition_state(info, STATE_FAILED)
             info.error = f"{type(exc).__name__}: {exc}"
             self._append_event(info, {"type": "final", "round": info.rounds,
                                       "reason": "error", "error": info.error})
@@ -455,14 +462,14 @@ class AgentService:
         info.prompt_tokens = int(result.get("prompt_tokens", 0))
         info.completion_tokens = int(result.get("completion_tokens", 0))
         if result.get("reason") == "error":
-            info.state = STATE_FAILED
+            self._transition_state(info, STATE_FAILED)
             info.error = result.get("error")
         elif result.get("reason") in ("token_cap", "max_rounds", "max_handoffs"):
             # 触顶属于护栏行为：结果仍交付，但标记部分完成信息
-            info.state = STATE_DONE
+            self._transition_state(info, STATE_DONE)
             info.error = result.get("error")
         else:
-            info.state = STATE_DONE
+            self._transition_state(info, STATE_DONE)
 
     # ---------- 两级模式（D7）：规划者 + 执行者 ----------
     async def run_dual(self, info: AgentRunInfo, planner_chat: Any, executor_chat: Any,