fix(pybot×actpred): v3 审计 P2 清零——验收链 fixed_range + 导出断言 + overrun 节流

- v3-P2-1: _npu_test/_onnx_validate 按 normalize_mode 分支（fixed_range 免 stats，
  HailoPredictor 透传模式，torch 参考同口径反归一化）
- v3-P2-2: export_deploy/export_onnx 一致性 diff 超 1e-4 即 sys.exit（P2-f 闭环）
- v3-P2-3: run_loop overrun 告警节流（第1次+每100次），降 hz 建议仅 >10% 输出
- P2-a1: default_column_weights 扩展列（action_dim>18）默认 1.0
- P2-c: 补 predictor 异常清理路径测试（异常传播 + finally 落盘）
- P2-d/P2-e: Resize 口径与 resume scheduler T_max 语义文档化
- R2-4: _set.sh 安装后断言 CPU 版 torch，误装 CUDA 即拦截
测试: pybot 89 passed (+1)、actpred 148 passed/18 skipped

pybot/pybot/robot.py

@@ -308,9 +308,13 @@ class RobotStruct:
                     time.sleep(remain)
                 else:
                     overrun += 1
-                    log.warning("loop overrun: step %d took %.1fms > %.1fms "
-                                "（推理+执行超过周期，实际频率已低于 %.1fHz）",
-                                step, (dt - remain) * 1000, dt * 1000, hz)
+                    # 告警节流：第 1 次及每 100 次汇总一次，避免持续 overrun
+                    # 刷屏（50% overrun @10Hz 时原实现每秒 5 条告警）
+                    if overrun == 1 or overrun % 100 == 0:
+                        log.warning("loop overrun ×%d: step %d took %.1fms > "
+                                    "%.1fms（推理+执行超过周期）",
+                                    overrun, step, (dt - remain) * 1000,
+                                    dt * 1000)
         finally:
             # max_steps 正常退出与 Ctrl-C（KeyboardInterrupt）都要落盘，
             # 否则闭环录制的 actions.npy/imu.npy 全部丢失
@@ -318,9 +322,14 @@ class RobotStruct:
                 path = recorder.flush()
                 log.info("episode saved: %s", path)
         if overrun:
-            log.warning("closed loop overrun %d/%d steps (%.0f%%)，"
-                        "建议降低 --hz 或检查推理后端延迟",
-                        overrun, step, 100.0 * overrun / max(step, 1))
+            ratio = overrun / max(step, 1)
+            if ratio > 0.10:    # 工程约束：overrun 比例 >10% 才建议降 hz
+                log.warning("closed loop overrun %d/%d steps (%.0f%%)，"
+                            "建议降低 --hz 或检查推理后端延迟",
+                            overrun, step, 100.0 * ratio)
+            else:
+                log.info("closed loop overrun %d/%d steps (%.1f%%)",
+                         overrun, step, 100.0 * ratio)
         log.info("closed loop stop after %d steps", step)
 
     def _loop_recent_frames(self, camera_var, F):

pybot/tests/test_loop.py

@@ -489,3 +489,27 @@ def test_robot_importable_without_numpy():
     r = subprocess.run([sys.executable, "-c", child],
                        capture_output=True, text=True, cwd=repo)
     assert r.returncode == 0 and "ok" in r.stdout, r.stderr
+
+
+def test_predictor_exception_still_flushes_recorder(mock_i2c, tmp_path):
+    """P2-c：predictor 中途抛异常时 run_loop 的 finally 仍落盘，
+    已录制 steps 不丢失且异常照常向上传播（由 with robot 清理硬件）"""
+    from pybot.recorder import EpisodeRecorder
+
+    class RaisingPredictor(MockPredictor):
+        def predict(self, images, action_history):
+            if self._step >= 1:
+                raise RuntimeError("backend boom")
+            return super().predict(images, action_history)
+
+    robot = _make_robot(LOOP_MT, mock_i2c)
+    robot.active()
+    pred = RaisingPredictor(frame_len=3, history_len=2, action_dim=4)
+    rec = EpisodeRecorder(str(tmp_path), action_dim=4)
+    with pytest.raises(RuntimeError, match="backend boom"):
+        robot.run_loop(pred, hz=1000.0, max_steps=5, recorder=rec)
+    robot.deactive()
+
+    # 第 1 步已录制（第 2 步 predict 抛异常），finally flush 已落盘
+    acts = np.load(tmp_path / "episode_0000" / "actions.npy")
+    assert acts.shape[0] == 1