exp(dagger-deploy): 轮次22——冻结编码器微调（开环保住15.8mm vs 19.5）但部署到达41.8%仍低于dag2 43.8%；训练侧四种方法均未超基线，训练探索到此平台

ml/train.py

@@ -73,6 +73,7 @@ def main():
     ap.add_argument('--config', default='ml/configs/base.json'); ap.add_argument('--steps', type=int, default=None)
     ap.add_argument('--resume', default=None); ap.add_argument('--no-eval', action='store_true')
     ap.add_argument('--init', default=None, help='只载入权重（优先 EMA）、优化器与步数从 0 开始（DAgger 微调用）')
+    ap.add_argument('--freeze-enc', action='store_true', help='冻结编码器（enc + enc_norm），只微调解码器/头')
     ap.add_argument('--lr', type=float, default=None); ap.add_argument('--warmup', type=int, default=None)
     ap.add_argument('--stop-at', type=int, default=0, help='自检用：跑到第 N 步按停止请求处理（存 last.pt、rc 143）')
     ap.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu'); ap.add_argument('--seed', type=int, default=0)
@@ -106,7 +107,7 @@ def main():
     print(f'train samples {data.n}, val tasks {len(val_tasks)}, act_scale {act_scale}, data_ver {data.data_ver}, device {dev}', flush=True)
     model = build(cfg).to(dev); model.act_scale = act_scale
     print(f'params {sum(p.numel() for p in model.parameters()) / 1e6:.2f} M', flush=True)
-    opt = torch.optim.AdamW(model.parameters(), lr=tc['lr'], betas=tuple(tc['betas']), weight_decay=tc['wd'])
+    opt = torch.optim.AdamW([p_ for p_ in model.parameters() if p_.requires_grad], lr=tc['lr'], betas=tuple(tc['betas']), weight_decay=tc['wd'])
     step0, best, bad_evals = 0, -1.0, 0
     def check_scale(ck, src):
         s0 = float(ck.get('act_scale', 1.0))