fix(closed_loop): 一致性打分错位一步（对齐 last[3]→last[4]，原来每次重规划奖励比上一计划慢一步的候选；Python 闭环与 ik-policy.js 同步修正）；rollout 加 select；diag_cl 删 F 段、加 G 段（1/4 段采样 × consist/value/first、4 段+安全层）；本机执行清单更新

ml/sim/closed_loop.py

@@ -65,8 +65,8 @@ def dls_steps(kin, q, T, R, n=3, scale=0.3, lam=0.5):
 
 
 @torch.no_grad()
-def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None, return_traj=False, use_sub=True, sub_override=None, qp0=None, steps=1, z_scale=1.0):
-    """use_sub=False：不用子目标（sf=0）；sub_override=(subs [B,2,6] 绝对关节角, need [B] bool)：冷启动时用给定途经点代替子目标头；qp0 [B,6]：起点的上一帧（默认 = q0，即静止起步）。steps：MeanFlow 采样段数；z_scale：噪声缩放（0 = 只取「均值」样本）。均为诊断用"""
+def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None, return_traj=False, use_sub=True, sub_override=None, qp0=None, steps=1, z_scale=1.0, select='consist'):
+    """use_sub=False：不用子目标（sf=0）；sub_override=(subs [B,2,6] 绝对关节角, need [B] bool)：冷启动时用给定途经点代替子目标头；qp0 [B,6]：起点的上一帧（默认 = q0，即静止起步）。steps：MeanFlow 采样段数；z_scale：噪声缩放（0 = 只取「均值」样本）；select：consist（部署口径：冷启动价值前瞻、之后一致性）/ value（每次都价值前瞻）/ first（不选，取第 0 条）。均为诊断用"""
     g = torch.Generator(device=device); g.manual_seed(seed)
     q0, T, R, ex, tr = tasks_to_tensors(tasks, device)
     B = q0.shape[0]
@@ -113,7 +113,9 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
                 if not use_sub: need = torch.zeros_like(need)
                 nsub = torch.where(need, torch.full_like(nsub, 2), torch.zeros_like(nsub)); sub_t.fill_(tnow)
             fut = qe[:, None, None] + torch.cumsum(inc[:, :, r:], 2)          # [B,K,H-r,6]
-            if last is None:
+            if select == 'first':
+                score = torch.zeros(B, K, device=device); score[:, 0] = 1
+            elif last is None or select == 'value':
                 # 短前瞻：候选末态的价值
                 qa = fut[:, :, -1].reshape(B * K, 6); qb = fut[:, :, -2].reshape(B * K, 6)
                 f2 = features(kin, qa, qb, T.repeat_interleave(K, 0), R.repeat_interleave(K, 0), torch.zeros(B * K, 3, 6, device=device),
@@ -121,9 +123,12 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
                 v2, _ = model.heads(model.encode(f2))
                 score = -v2.view(B, K)
             else:
-                n = min(fut.shape[2], last.shape[1] - 3)
+                # 2026-10-19 修错位：last[i] = 上一计划从 qe_old 走 i 步后的位置；提交 3 步后 qe_new = last[3]，
+                # 新候选 fut[0] 是 qe_new 再走 1 步 = 对应 last[4]。原来对齐 last[3]，等于奖励「比上一计划慢一步」的候选，
+                # 每次重规划都挑最拖的那条，闭环越走越慢。JS 部署（ik-policy.js）同一处同样修正。
+                n = min(fut.shape[2], last.shape[1] - 4)
                 w = 0.85 ** torch.arange(n, device=device, dtype=torch.float32)
-                d2 = (((fut[:, :, :n] - last[:, None, 3:3 + n]) / step) ** 2).sum(-1)
+                d2 = (((fut[:, :, :n] - last[:, None, 4:4 + n]) / step) ** 2).sum(-1)
                 score = -(d2 * w).sum(-1)
             order = score.argsort(-1, descending=True)
             pick = order[:, 0]