fix(kin/diag): gaps 默认不拼胶囊对（pairs=True 仅 APG 用；实测 4096×40 峰值 4.46 GiB 全在拼接 8.6e7 对）；diag_cl 删 D 段（子目标已排除）、加 E 段（开环幅度与抹 dq 敏感度、老师中段带速度/静止起步闭环）；rollout 加 qp0；本机执行清单更新

ml/common/kin_torch.py

@@ -129,24 +129,27 @@ class Kin:
 
     @staticmethod
     def _cat_gaps(parts, starts):
-        return {'gm': torch.cat([p['gm'] for p in parts], 0),
-                'b': torch.cat([p['b'] + s for p, s in zip(parts, starts)]),
-                'i': torch.cat([p['i'] for p in parts]), 'j': torch.cat([p['j'] for p in parts]),
-                'gap': torch.cat([p['gap'] for p in parts])}
+        out = {'gm': torch.cat([p.pop('gm') for p in parts], 0)}
+        if 'b' in parts[0]:
+            for p, s in zip(parts, starts): p['b'] += s
+            for k in ('b', 'i', 'j', 'gap'): out[k] = torch.cat([p.pop(k) for p in parts])   # 逐键拼、拼完即释放
+        return out
 
-    def gaps(self, q, cap=40.0, xf=None, max_pairs=None):
+    def gaps(self, q, cap=40.0, xf=None, max_pairs=None, pairs=False):
         """返回 dict：
              gm  [B,49]：组对最小间隙 min(gap, cap)（gA*7+gB，gA<gB；其余为 cap）
-             b, i, j, gap：幸存胶囊对（gap 可微）
+             pairs=True 时另有 b, i, j, gap：幸存胶囊对（gap 可微；只有 APG 的 hinge 与 check_parity 需要）
+           2026-10-19（用户实测 4096 点 cap 40 峰值 4.46 GiB 全在拼接 8.6e7 对的 b/i/j/gap）：
+             默认只返回 gm，每块算完就丢掉胶囊对，峰值只剩单块。
         """
         xf = xf if xf is not None else self.cum_xf(q)
         Rs, Ts = xf
         B = q.shape[0]
         step = max(1, self.COARSE_ENTRIES // int(self.cp_a.shape[0]))
         if B <= step:
-            return self._gaps_split(q, cap, xf)
+            return self._gaps_split(q, cap, xf, pairs)
         st = list(range(0, B, step))
-        return self._cat_gaps([self._gaps_split(q[s:s + step], cap, (Rs[s:s + step], Ts[s:s + step])) for s in st], st)
+        return self._cat_gaps([self._gaps_split(q[s:s + step], cap, (Rs[s:s + step], Ts[s:s + step]), pairs) for s in st], st)
 
     def _coarse(self, q, cap, xf):
         Rs, Ts = xf
@@ -155,7 +158,7 @@ class Kin:
         del cc
         return d2 <= (self.cp_lim + cap).pow(2)
 
-    def _gaps_split(self, q, cap, xf):
+    def _gaps_split(self, q, cap, xf, pairs):
         """粗筛已在显存预算内：按幸存细对数再切块做精算"""
         Rs, Ts = xf
         B = q.shape[0]
@@ -171,10 +174,10 @@ class Kin:
                 bounds.append(B)
             del per
         if len(bounds) == 2:
-            return self._gaps_one(q, cap, xf)
-        return self._cat_gaps([self._gaps_one(q[s:e], cap, (Rs[s:e], Ts[s:e])) for s, e in zip(bounds[:-1], bounds[1:])], bounds[:-1])
+            return self._gaps_one(q, cap, xf, pairs)
+        return self._cat_gaps([self._gaps_one(q[s:e], cap, (Rs[s:e], Ts[s:e]), pairs) for s, e in zip(bounds[:-1], bounds[1:])], bounds[:-1])
 
-    def _gaps_one(self, q, cap, xf):
+    def _gaps_one(self, q, cap, xf, pairs):
         Rs, Ts = xf
         B = q.shape[0]
         with torch.no_grad():
@@ -195,6 +198,7 @@ class Kin:
         gap = d - self.pair_need[pidx].to(q.dtype)
         gm = torch.full((B * 49,), float(cap), dtype=q.dtype, device=q.device)
         gm = gm.scatter_reduce(0, rep_b * 49 + self.pair_gp[pidx], gap.clamp(max=cap), reduce='amin', include_self=True)
+        if not pairs: return {'gm': gm.view(B, 49)}
         return {'gm': gm.view(B, 49), 'b': rep_b, 'i': pi, 'j': pj, 'gap': gap}
 
     def min_gap(self, q, cap=40.0):