feat+exp: 轮次39——候选档位开关 --z_scale 与行程惩罚 select=ep_len/--len_w（零权重与 endpoint 逐位相同，2 条机制单测）

ml/eval/eval_cl.py

@@ -94,6 +94,8 @@ def main():
     ap.add_argument('--gap_margin', type=float, default=0.0, help='安全层要求的间隙余量 mm（>0 时宽相位 cap 自动抬到 margin+0.5）')
     ap.add_argument('--cert_batch', type=int, default=0, help='1=一次 gaps 调用覆盖全部候选（压最坏情况，典型路径略贵）')
     ap.add_argument('--steps', type=int, default=1, help='MeanFlow 采样段数；1=线上一口径（历史发布数字全是 1）')
+    ap.add_argument('--z_scale', type=float, default=1.0, help='候选噪声缩放：>1 让 K 条候选更分散（测"全拒是没有备选"还是"真的没路"）')
+    ap.add_argument('--len_w', type=float, default=0.0, help="select=ep_len 时行程惩罚权重（mm / 每轴归一 STEP 的累计路程）")
     ap.add_argument('--backoff', type=int, default=0, help='全拒时把首选候选按 1/2、1/4…缩步重试的档数（需 --cert_batch 1）')
     ap.add_argument('--max_sec', type=float, default=None)
     ap.add_argument('--seed', type=int, default=0)
@@ -117,11 +119,11 @@ def main():
         'kin_fp': kin.fp, 'act_scale': float(model.act_scale),
     }
     t0 = time.perf_counter()
-    r, pers = run(model, kin, tasks, a.device, K=a.K, safety=bool(a.safety), seed=a.seed, max_sec=a.max_sec, steps=a.steps,
+    r, pers = run(model, kin, tasks, a.device, K=a.K, safety=bool(a.safety), seed=a.seed, max_sec=a.max_sec, steps=a.steps, z_scale=a.z_scale,
             safety_partial=(a.commit if (a.safety and a.partial) else 0), commit_n=a.commit,
             select=a.select, replan=a.replan, fb_gain=a.fb, fb_scale=a.fb_scale, fb_near=a.fb_near, fb_fallback=a.fb_fallback, fb_ki=a.ki, fb_decay=a.kdecay, fb_imax=a.kimax,
             fb_every=a.fb_every, safety_try=a.safety_try, sel_w=a.sel_w, dls_safety=a.dls_safety, fb_clear=a.fb_clear, fb_cmin=a.fb_cmin, sub_refresh=a.sub_refresh, fb_block_gain=a.fb_block_gain,
-            inc_cap=a.inc_cap, gap_margin=a.gap_margin, cert_batch=a.cert_batch, backoff=a.backoff, return_traj=bool(a.trace))
+            inc_cap=a.inc_cap, gap_margin=a.gap_margin, cert_batch=a.cert_batch, backoff=a.backoff, len_w=a.len_w, return_traj=bool(a.trace))
     wall = time.perf_counter() - t0
     result = {**meta, 'metrics': r, 'simulation_wall_s': wall}
     print(json.dumps(json_finite(result), ensure_ascii=False, allow_nan=False), flush=True)