exp(dagger-deploy): 轮次21——DAgger-3（部署配置采集500题+15000步微调）负结论：部署口径 39.5% vs dag2 43.8%，开环退化到19.5mm；下一步试冻结编码器只调解码器

ml/configs/base_dag3.json

@@ -0,0 +1,11 @@
+{
+  "model": {"d": 192, "heads": 8, "ff": 768, "enc": 4, "dec": 4},
+  "train": {
+    "batch": 1024, "lr": 1e-3, "warmup": 1000, "steps": 24000, "wd": 0.01, "betas": [0.9, 0.95],
+    "amp": "bf16", "grad_clip": 1.0,
+    "w_mf": 1.0, "w_fk": 0.5, "w_smooth": 0.05, "w_value": 0.1, "w_sub": 0.1,
+    "p_neq": 0.5, "p_edge": 0.1, "p_neq_warm": 5000, "fk_tmax": 0.3, "fk_caps": 96,
+    "ema": 0.999, "aug_pre_drop": 0.15, "aug_pre_sigma": 0.1, "aug_dq_sigma": 0.3,
+    "ckpt_min": 10, "eval_every": 2000, "eval_tasks": 512, "log_every": 50, "patience_evals": 5
+  }
+}

ml/dagger/collect.py

@@ -30,6 +30,8 @@ def main():
     ap.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu')
     ap.add_argument('--fb_gain', type=float, default=0.0, help='部署分布采集：目标反馈增益（与 rollout 一致）')
     ap.add_argument('--safety', type=int, default=0, help='部署分布采集：是否开安全层')
+    ap.add_argument('--replan', default='r3'); ap.add_argument('--commit_n', type=int, default=3)
+    ap.add_argument('--safety_partial', type=int, default=0)
     a = ap.parse_args()
     gap_min = float(os.environ.get('MIN_GAP_MM') or 3)
     os.makedirs(a.out, exist_ok=True)
@@ -43,7 +45,8 @@ def main():
     for i in range(0, len(tasks), a.bs):
         tb = tasks[i:i + a.bs]
         with torch.no_grad():
-            _, per = rollout(model, kin, tb, a.device, K=8, safety=bool(a.safety), seed=a.seed * 100003 + i, return_traj=True, fb_gain=a.fb_gain)
+            _, per = rollout(model, kin, tb, a.device, K=8, safety=bool(a.safety), seed=a.seed * 100003 + i, return_traj=True,
+                            fb_gain=a.fb_gain, replan=a.replan, commit_n=a.commit_n, safety_partial=a.safety_partial)
         tr = per['traj'].to(a.device)                                   # [B,T,6]
         B, Tn, _ = tr.shape
         _, T, R, _, _ = tasks_to_tensors(tb, a.device)

ml/dagger/run.sh

@@ -24,7 +24,7 @@ for ((r=1; r<=R; r++)); do
   SAMPLES="$SAMPLES $SD"
   if [ -f $OUT/done ]; then echo "[dagger] 第 $r 轮已完成，跳过"; CK=$($PY -m ml.common.ckpt pick $OUT); continue; fi
   if [ -f $D/collect.json ]; then echo "[dagger] 第 $r 轮起点已采集，跳过"
-  else echo "[dagger] 第 $r 轮：用 $CK 闭环采起点（$N 题）"; $PY -m ml.dagger.collect --ckpt $CK --samples ml/data/samples --n $N --seed $r --parts $P --out $D --fb_gain ${FB_GAIN:-0} --safety ${DAG_SAFETY:-0}; fi
+  else echo "[dagger] 第 $r 轮：用 $CK 闭环采起点（$N 题）"; $PY -m ml.dagger.collect --ckpt $CK --samples ml/data/samples --n $N --seed $r --parts $P --out $D --fb_gain ${FB_GAIN:-0} --safety ${DAG_SAFETY:-0} --replan ${DAG_REPLAN:-r3} --commit_n ${DAG_COMMIT_N:-3} --safety_partial ${DAG_PARTIAL:-0}; fi
   cat $D/collect.json
   pids=()
   for f in $D/tasks_*.jsonl; do