exp+fix: 候选塌缩定案(steps=2 把K条候选抹平35倍)+z_scale反例；补 n_plans/brake_rate 核算§9.2刹停率；记录我在链式作业排队期间改库导致两臂未跑的事故

ml/eval/eval_cl.py

@@ -54,6 +54,9 @@ def run(model, kin, tasks, device, bs=256, **kw):
         out.update({'reverse_p50': median(rr), 'reverse_mean': rr.mean().item(),
                     'reverse_moving_p50': median(flat['reverse_rate_moving']),   # 只在动了的步上算，不被刹停稀释
                     'moving_share_p50': median(flat['moving_share'])})          # 活动窗内动了的步占比（=1−站住比例）
+    if 'brake_rate' in flat:   # §9.2 判据是"刹停 <2% 推理周期"，所以要按规划次数归一，不能只报绝对次数
+        out.update({'brake_rate_mean': flat['brake_rate'].mean().item(), 'brake_rate_p50': median(flat['brake_rate']),
+                    'n_plans_mean': flat['n_plans'].mean().item()})
     if 'backoff' in flat and flat['backoff'].any():   # backoff=缩步顶上的次数；brakes 现在只记连最小档都不安全的
         out['backoff_mean'] = flat['backoff'].mean().item()      # ⇒ 原来的"全拒"事件数 = 两者之和
         out['reject_all_mean'] = (flat['backoff'] + flat['brakes']).mean().item()
@@ -139,7 +142,7 @@ def main():
         Path(a.out).write_text(json.dumps(json_finite(result), indent=2, ensure_ascii=False, allow_nan=False) + '\n', encoding='utf-8')
     if a.dump:
         rows = []
-        keys = ['done', 'ok', 't_done', 'limit', 'col', 'brakes', 'err', 'first_hit', 'dls_it', 'goal_ratio_2s']
+        keys = ['done', 'ok', 't_done', 'limit', 'col', 'brakes', 'err', 'first_hit', 'dls_it', 'goal_ratio_2s', 'brake_rate', 'n_plans']
         flat = {k: torch.cat([p[k] for _, p, _ in pers]).tolist() for k in keys}
         for i, t in enumerate(tasks):
             rows.append({**t, 'i': i, **{k: flat[k][i] for k in keys}})