fix(actpred): 修复 Hailo HEF 编译链路的输入断连与封装问题

- hailo_infer: 帧输入按 input_layerN 排序绑定（vstream 报告顺序实测
  为 1,5,4,3,2，按报告顺序会 frame1/frame2 错位），新增 --pred-token
- gen_calibration: pred_token 校准改为学习值 ±30% 抖动，修复常量
  校准量化退化（scale=18571.7，任何输入都被量化成 0）
- export_deploy: 新增 --pred-token-out 保存学习值供推理端喂入
- dfc_compile: parser/optimize 后自动检查输入层连通性，断开即报错
- 新增 _npu_test.py（NPU 4 角敏感性复验）与 _onnx_validate.py
  （编译前 ONNX 数值门槛），test_actpred 补乱序帧序用例

actpred/_onnx_validate.py

@@ -0,0 +1,92 @@
+"""部署版 ONNX 数值验证：输入敏感性 + 与 PyTorch 一致性（重编译前门槛）
+
+用法: venv/bin/python _onnx_validate.py
+"""
+import sys
+sys.path.insert(0, "/opt/blackbird/actpred")
+
+import numpy as np
+import torch
+from PIL import Image
+from torchvision import transforms
+import onnxruntime as ort
+
+from config import parse_checkpoint_config
+from model import ActionPredictor
+
+CKPT = "checkpoints/pretrain_best_inference.pt"
+ONNX = "checkpoints/actpred_deploy.onnx"
+IMG_DIR = "data/episode_0000/images"
+
+ck = torch.load(CKPT, map_location="cpu", weights_only=True)
+cfg = parse_checkpoint_config(ck["config"])
+model = ActionPredictor(cfg.model, cfg.data, pretrained=False)
+model.load_state_dict(ck["model_state_dict"])
+model.eval()
+
+sess = ort.InferenceSession(ONNX)
+in_names = [i.name for i in sess.get_inputs()]
+print("ONNX inputs:", in_names)
+
+h, w = cfg.data.image_height, cfg.data.image_width
+tf = transforms.Compose([
+    transforms.Resize((h, w)),
+    transforms.ToTensor(),
+    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+])
+
+imgs_real = [np.array(Image.open(f"{IMG_DIR}/000{i}.jpg").convert("RGB"))
+             for i in range(3)]
+imgs_zero = [np.zeros_like(imgs_real[0]) for _ in range(3)]
+
+hist_real = np.array([
+    [45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 10, 200, 0, 0, 0, 0, 0, 0],
+    [45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 20, 0, 0, 0, 0, 0, 0],
+    [45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 0, 0, 0, 0, 0, 0, 0],
+    [45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 0, 0, 0, 0, 0, 0, 0],
+    [45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 0, 0, 0, 0, 0, 0, 0],
+], dtype=np.float32)
+hist_zero = np.zeros((5, 18), dtype=np.float32)
+
+pred_learned = model.pred_token.data.unsqueeze(1).numpy()   # (1,1,1,128)
+pred_none = np.zeros((1, 1, 1, 128), dtype=np.float32)
+
+
+def frames_nhwc(imgs):
+    """列表帧 → (1,H,W,3) 归一化 float32 NHWC × frame_len"""
+    chw = np.stack([tf(Image.fromarray(im)).numpy() for im in imgs])
+    return [chw[i].transpose(1, 2, 0)[None] for i in range(len(imgs))]
+
+
+def run_onnx(imgs, hist, pred):
+    feeds = {f"frame{i}": f for i, f in enumerate(frames_nhwc(imgs))}
+    feeds["action_history"] = hist[None, None]
+    feeds["pred_token"] = pred
+    return sess.run(None, feeds)[0].reshape(-1)
+
+
+def run_torch(imgs, hist):
+    x = torch.stack([tf(Image.fromarray(im)) for im in imgs]).unsqueeze(0)
+    hh = torch.tensor(hist).unsqueeze(0)
+    with torch.no_grad():
+        return model(x, hh).squeeze(0).numpy()
+
+
+a = run_onnx(imgs_real, hist_real, pred_learned)
+b = run_onnx(imgs_zero, hist_real, pred_learned)
+c = run_onnx(imgs_real, hist_zero, pred_learned)
+d = run_onnx(imgs_zero, hist_zero, pred_learned)
+p0 = run_onnx(imgs_real, hist_real, pred_none)
+
+print(f"onnx real_img+real_hist: {np.round(a[:6], 3)}...")
+print(f"onnx pred_token=0 diff : max={np.abs(a - p0).max():.4f}")
+print(f"img sensitivity  (a-b): max={np.abs(a - b).max():.4f}")
+print(f"hist sensitivity (a-c): max={np.abs(a - c).max():.4f}")
+
+ref = run_torch(imgs_real, hist_real)
+print(f"torch reference        : {np.round(ref[:6], 3)}...")
+print(f"onnx vs torch max_diff : {np.abs(a - ref).max():.6e}")
+
+ok = (np.abs(a - b).max() > 1e-4 and np.abs(a - c).max() > 1e-4
+      and np.abs(a - ref).max() < 1e-4)
+print("PASS" if ok else "FAIL")