fix(kin/diag): kin.gaps 粗筛段也按点分块（COARSE_ENTRIES=3e7，≤446 点/块；实测上一版粗筛 4096 点仍 9.3 GiB）；diag_cl 加 D 段（子目标头精度 / 不用子目标 / 老师真实途经点）与目标误差进度比；rollout 加 use_sub、sub_override；本机执行清单更新

ml/eval/diag_cl.py

@@ -27,6 +27,9 @@ GROUPS = [('sincos', 0, 12), ('joint_norm', 12, 15), ('dq', 15, 21), ('p', 21, 2
           ('sub', 68, 74), ('sub_flag', 74, 75), ('gaps', 75, 83)]
 
 
+PROG_T = (0.5, 1.0, 2.0)     # 目标误差 / 起点误差 的取样时刻（s）
+
+
 def pct(v, p):
     v = sorted(v)
     return v[min(len(v) - 1, int(len(v) * p))] if v else None
@@ -81,6 +84,7 @@ def main():
 
     # ---------- C：闭环失败形态 ----------
     first_t, col_steps, pairs, min_err, reach = [], [], {}, [], 0
+    prog = {ts: [] for ts in PROG_T}
     reach_s, brakes_s, col_s = 0, [], 0
     for i in range(0, len(tasks), a.bs):
         ch = tasks[i:i + a.bs]
@@ -95,6 +99,9 @@ def main():
             Rt = torch.tensor([t['R'] if t.get('R') else [0.] * 9 for t in ch], dtype=torch.float32, device=dev)
             e, _ = goal_err(kin, tr.reshape(-1, 6), Tt.repeat_interleave(Tn, 0), Rt.repeat_interleave(Tn, 0))
             e = e.view(B, Tn)
+            for b in range(B):
+                for ts in PROG_T:
+                    k = min(Tn - 1, int(round(ts / DT))); prog[ts].append((e[b, k] / e[b, 0].clamp(min=1e-3)).item())
             # 到达后轨迹被冻结（d=0），之后不再计
             for b in range(B):
                 if per['done'][b]: reach += 1
@@ -113,6 +120,38 @@ def main():
                 'first_hit_pair(group a-b)': dict(sorted(pairs.items(), key=lambda kv: -kv[1])[:8]),
                 'min_goal_err_mm': {'p10': pct(min_err, .1), 'p50': pct(min_err, .5), 'p90': pct(min_err, .9)},
                 'safety_on': {'reach': reach_s / n, 'collided': col_s / n, 'brakes_p50': pct(brakes_s, .5)}}
+    out['C']['goal_err_ratio_p50'] = {'%.1fs' % ts: pct(v, .5) for ts, v in prog.items()}   # <1 = 在靠近目标
+
+    # ---------- D：子目标是不是元凶（2026-10-19 加：C 显示全程最小目标误差中位 254 mm，策略根本没朝目标走） ----------
+    # 训练时特征里的子目标是老师真实途经点（80% 概率给），闭环里却换成子目标头自己的预测；预测偏了，策略就会忠实地朝错的方向走。
+    sg_by_tid = {}
+    for i, t in zip(idx, tk):
+        qrow = torch.atan2(vd.X[i, 0:12:2].float(), vd.X[i, 1:12:2].float()) * 180 / torch.pi
+        sgr = vd.SG[i].float()
+        sg_by_tid[t['tid']] = (qrow[None] + sgr[:12].view(2, 6) * 180, bool(sgr[12] > 0.5))
+    # D0：冷启动子目标头精度（起点特征里抹掉数据集的真子目标，与闭环冷启动同口径）
+    Xs = X0.clone(); Xs[:, 68:75] = 0
+    with torch.no_grad(): _, sp = model.heads(model.encode(Xs))
+    lab = torch.stack([vd.SG[i].float() for i in idx])
+    need_l = lab[:, 12] > 0.5; need_p = sp[:, 12] > 0.5
+    via_err = ((sp[:, :6] - lab[:, :6]) * 180).abs().max(-1).values
+    out['D0_subhead'] = {'n': len(idx), 'need_label_rate': need_l.float().mean().item(), 'need_acc': (need_l == need_p).float().mean().item(),
+                         'via1_maxjoint_err_deg_p50': pct(via_err[need_l].tolist(), .5), 'via1_maxjoint_err_deg_p90': pct(via_err[need_l].tolist(), .9)}
+    # D1 不用子目标、D2 用老师真实途经点（oracle），同一批题、同一随机种子，对比 C 的基线
+    def run(tag, **kw):
+        rc = cc_ = 0; me = []
+        for i in range(0, len(tasks), a.bs):
+            ch = tasks[i:i + a.bs]; extra = dict(kw)
+            if kw.get('sub_override') == 'oracle':
+                subs = torch.stack([sg_by_tid[t['tid']][0] if t['tid'] in sg_by_tid else torch.zeros(2, 6, device=dev) for t in ch])
+                need = torch.tensor([sg_by_tid[t['tid']][1] if t['tid'] in sg_by_tid else False for t in ch], device=dev)
+                extra['sub_override'] = (subs, need)
+            with torch.no_grad():
+                _, pp = rollout(model, kin, ch, dev, K=8, safety=False, seed=i, **extra)
+            rc += int(pp['done'].sum()); cc_ += int((pp['col'] > 0).sum()); me += pp['err'].tolist()
+        out[tag] = {'reach': rc / n, 'collided': cc_ / n, 'end_goal_err_mm_p50': pct(me, .5)}
+    run('D1_no_sub', use_sub=False)
+    run('D2_oracle_sub', sub_override='oracle')
     s = json.dumps(out, indent=1, ensure_ascii=False); print(s)
     os.makedirs(os.path.dirname(a.out) or '.', exist_ok=True)
     open(a.out, 'w', encoding='utf-8').write(s)