exp: 补提交轮次40 的 K 成本探针产物（此前漏在未跟踪文件里，是候选档位成本-收益曲线的一手数据）

之前的"倒序"只在 ## 轮次 NN 之间成立：记分板/续篇/按 issue 号命名的条目(如 #27、#24、判别定案)
各有自己的位置，所以打开后第 4 屏才看到新轮次，且 轮次 39 那一族曾压在 52 上面。
现在规则统一：文件头 + 轮次统计 + 算法优化候选常驻最前，其后所有小节按轮次号降序，
无轮次号的行政/教训类(分支清理、事故记录)按约定放最后。

轮次号取自标题里的"轮次 NN"；标题没有的从正文找(#27 那种)，正文也没有的走显式定年表
(两条 #21 的条目按 round40 产物定为轮次 40)。同轮次内保持原有相对顺序，不乱改。

脚本 ml/runs/_sort_log_reverse.py（默认 dry-run，--apply 才写盘），带三重守恒断言：
字符数相同、逐行多重集相同、排序前后 diff -q 与 HEAD 完全一致 ⇒ 这次改动是纯重排
(72 增/72 删只是块边界)。LF 未被整体改写成 CRLF。

docs/experiment_round29.json

@@ -0,0 +1,62 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 512,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round29_rows.json",
+    "out": "docs/experiment_round29.json",
+    "trace": "ml/runs/round29_trace.pt",
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "1fd72f71eaf8170b45782beede0d91544e1f7142e98c0642560af814382f01ef",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 512,
+    "n_ref": 512,
+    "success": 0.55078125,
+    "collision_free_success": 0.546875,
+    "time_ok_1p3": 0.361328125,
+    "collision_rate": 0.017578125,
+    "collision_steps_mean": 0.1015625,
+    "travel_ratio_median": 2.425084352493286,
+    "time_mean_s": 2.322056770324707,
+    "brakes_mean": 19.228515625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.25053170323371887,
+    "err_fail_med": 194.51763916015625,
+    "err_success_med": 0.1961677223443985,
+    "dls_share": 0.634765625,
+    "time_budget": "teacher×2(≥3 s)"
+  },
+  "simulation_wall_s": 488.4280448999925
+}

docs/experiment_round29_collisions.json

@@ -0,0 +1,779 @@
+{
+  "trace": "ml/runs/round29_trace.pt",
+  "n": 512,
+  "reported_collision_tasks": 9,
+  "replayed_first_hits32": 9,
+  "replayed_first_hits64": 9,
+  "scope": "Replay reported first hits only; same geometry kernel at float32/float64, not a full safety audit.",
+  "rows": [
+    {
+      "i": 21,
+      "task": {
+        "tid": 1223280991,
+        "id": "s0001_000564",
+        "cat": "A",
+        "mode": "pose",
+        "q0": [
+          178.767,
+          101.917,
+          25.283,
+          31.588,
+          -12.179,
+          -6.464
+        ],
+        "T": [
+          320.779,
+          79.294,
+          350.853
+        ],
+        "R": [
+          -0.5548626629953249,
+          -0.1382477148217311,
+          -0.820374910976869,
+          -0.29956278115972873,
+          -0.8867511636570162,
+          0.35204334093514156,
+          -0.7761375943462505,
+          0.44108949556693955,
+          0.4506112421399604
+        ],
+        "exec_s": 2.86,
+        "travel": 581.023,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 5.659999847412109,
+      "dls_it": -1,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.5,
+        0.35770511627197266,
+        -0.3957242965698242
+      ],
+      "gap64_mm": [
+        0.5,
+        0.5,
+        0.3577231796451059,
+        -0.39572636491366353
+      ],
+      "pair32": [
+        0,
+        0,
+        5,
+        5
+      ],
+      "pair64": [
+        0,
+        0,
+        5,
+        5
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        139.78468322753906,
+        87.256591796875,
+        6.928672790527344,
+        103.78778839111328,
+        -94.0,
+        19.4688720703125
+      ],
+      "endpoint": [
+        139.50450134277344,
+        87.4175796508789,
+        6.71871280670166,
+        104.29376983642578,
+        -94.0,
+        21.218488693237305
+      ],
+      "dense_nsegments": 175,
+      "dense_gap64_min_mm": -0.39572636491366353
+    },
+    {
+      "i": 84,
+      "task": {
+        "tid": 1268963905,
+        "id": "s0002_000538",
+        "cat": "B",
+        "mode": "pose",
+        "q0": [
+          -49.198,
+          52.395,
+          108.796,
+          75.534,
+          82.302,
+          -121.823
+        ],
+        "T": [
+          -85.062,
+          275.378,
+          888.929
+        ],
+        "R": [
+          0.5139412462799992,
+          -0.4285630099475821,
+          -0.7431003578769356,
+          -0.5533754886400825,
+          -0.8275489878032067,
+          0.09454227286372051,
+          -0.6554692700230204,
+          0.36262435010704513,
+          -0.6624678231921397
+        ],
+        "exec_s": 3.2600000000000002,
+        "travel": 349.09,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 5.980000019073486,
+      "dls_it": -1,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.030836105346679688,
+        -0.1565418243408203,
+        -0.11790847778320312
+      ],
+      "gap64_mm": [
+        0.5,
+        0.030838235678215753,
+        -0.15655898804830848,
+        -0.11791701257236653
+      ],
+      "pair32": [
+        0,
+        6,
+        6,
+        6
+      ],
+      "pair64": [
+        0,
+        6,
+        6,
+        6
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        -69.87329864501953,
+        36.559112548828125,
+        2.094588279724121,
+        102.22724151611328,
+        89.0,
+        -212.1793212890625
+      ],
+      "endpoint": [
+        -69.79582214355469,
+        37.351837158203125,
+        -0.033895254135131836,
+        102.45509338378906,
+        89.0,
+        -214.75440979003906
+      ],
+      "dense_nsegments": 258,
+      "dense_gap64_min_mm": -0.16882067924099076
+    },
+    {
+      "i": 145,
+      "task": {
+        "tid": 593213945,
+        "id": "s0003_000414",
+        "cat": "B",
+        "mode": "pose",
+        "q0": [
+          -49.395,
+          84.783,
+          2.744,
+          14.841,
+          -86.783,
+          81.932
+        ],
+        "T": [
+          5.391,
+          506.723,
+          383.367
+        ],
+        "R": [
+          -0.5269859333114162,
+          -0.7554822816510317,
+          -0.38927156099983334,
+          0.6402732880654818,
+          -0.05172298162073713,
+          -0.7664038424760652,
+          0.5588702377848656,
+          -0.6531242265324508,
+          0.5109724082909718
+        ],
+        "exec_s": 2.96,
+        "travel": 560.55,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 5.599999904632568,
+      "dls_it": -1,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.5,
+        0.3180522918701172,
+        -0.19827747344970703
+      ],
+      "gap64_mm": [
+        0.5,
+        0.5,
+        0.318071359587325,
+        -0.19825289671655533
+      ],
+      "pair32": [
+        0,
+        0,
+        5,
+        5
+      ],
+      "pair64": [
+        0,
+        0,
+        5,
+        5
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        -36.05080032348633,
+        49.852294921875,
+        22.054914474487305,
+        67.3380126953125,
+        -94.0,
+        79.87274932861328
+      ],
+      "endpoint": [
+        -35.60316848754883,
+        48.9223747253418,
+        21.386226654052734,
+        67.87915802001953,
+        -94.0,
+        79.27398681640625
+      ],
+      "dense_nsegments": 93,
+      "dense_gap64_min_mm": -0.19825289671655533
+    },
+    {
+      "i": 164,
+      "task": {
+        "tid": 1743026973,
+        "id": "s0003_000814",
+        "cat": "C",
+        "mode": "pos",
+        "q0": [
+          -65.982,
+          5.595,
+          -64.669,
+          -72.912,
+          85.732,
+          -86.377
+        ],
+        "T": [
+          -189.354,
+          221.034,
+          302.457
+        ],
+        "R": null,
+        "exec_s": 2.3000000000000003,
+        "travel": 306.824,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 3.0399999618530273,
+      "dls_it": -1,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.5,
+        0.5,
+        -0.12672901153564453
+      ],
+      "gap64_mm": [
+        0.5,
+        0.5,
+        0.5,
+        -0.12673094091899273
+      ],
+      "pair32": [
+        0,
+        0,
+        0,
+        5
+      ],
+      "pair64": [
+        0,
+        0,
+        0,
+        5
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        -33.6672248840332,
+        39.63850021362305,
+        -60.174774169921875,
+        -108.2870864868164,
+        88.27082061767578,
+        -74.56200408935547
+      ],
+      "endpoint": [
+        -32.8936882019043,
+        39.478729248046875,
+        -58.639915466308594,
+        -107.65913391113281,
+        89.0,
+        -74.92655944824219
+      ],
+      "dense_nsegments": 154,
+      "dense_gap64_min_mm": -0.12673094091899273
+    },
+    {
+      "i": 175,
+      "task": {
+        "tid": 444406422,
+        "id": "s0003_001100",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          -28.854,
+          43.33,
+          -137.874,
+          -6.895,
+          -88.001,
+          -43.745
+        ],
+        "T": [
+          379.105,
+          293.88,
+          653.541
+        ],
+        "R": null,
+        "exec_s": 2.58,
+        "travel": 465.057,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 0.8399999737739563,
+      "dls_it": 144,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.4698514938354492,
+        -0.6612262725830078,
+        -0.9359521865844727
+      ],
+      "gap64_mm": [
+        0.5,
+        0.4698356780126254,
+        -0.6612412101671552,
+        -0.9359581587234
+      ],
+      "pair32": [
+        0,
+        6,
+        6,
+        6
+      ],
+      "pair64": [
+        0,
+        6,
+        6,
+        6
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        -42.48794937133789,
+        11.756612777709961,
+        -124.7530288696289,
+        -46.12489700317383,
+        -94.0,
+        -54.57075500488281
+      ],
+      "endpoint": [
+        -42.73639678955078,
+        10.6148681640625,
+        -124.47537231445312,
+        -48.52891159057617,
+        -94.0,
+        -55.451263427734375
+      ],
+      "dense_nsegments": 241,
+      "dense_gap64_min_mm": -0.9544276230224238
+    },
+    {
+      "i": 180,
+      "task": {
+        "tid": 737760595,
+        "id": "s0003_001280",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          41.213,
+          49.353,
+          -149.808,
+          -8.981,
+          -88.566,
+          83.45
+        ],
+        "T": [
+          118.145,
+          2.733,
+          266.082
+        ],
+        "R": null,
+        "exec_s": 1.78,
+        "travel": 228.67,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 1.1399999856948853,
+      "dls_it": 102,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.06517887115478516,
+        -0.0172576904296875,
+        0.013133049011230469
+      ],
+      "gap64_mm": [
+        0.5,
+        0.06521150921177998,
+        -0.017227151210720493,
+        0.013145878955246104
+      ],
+      "pair32": [
+        0,
+        5,
+        5,
+        5
+      ],
+      "pair64": [
+        0,
+        5,
+        5,
+        5
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        1.8212896585464478,
+        -9.327469825744629,
+        -150.87704467773438,
+        -23.00735855102539,
+        -94.0,
+        93.73664855957031
+      ],
+      "endpoint": [
+        1.3598560094833374,
+        -9.66404914855957,
+        -150.52992248535156,
+        -23.200042724609375,
+        -94.0,
+        92.69589233398438
+      ],
+      "dense_nsegments": 105,
+      "dense_gap64_min_mm": -0.020246891242656773
+    },
+    {
+      "i": 215,
+      "task": {
+        "tid": 169730876,
+        "id": "s0004_000544",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          47.051,
+          -1.179,
+          -94.31,
+          -43.128,
+          67.067,
+          -119.867
+        ],
+        "T": [
+          -142.696,
+          -139.159,
+          726.163
+        ],
+        "R": null,
+        "exec_s": 2.16,
+        "travel": 461.684,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 3.5399999618530273,
+      "dls_it": -1,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.5,
+        0.5,
+        -0.08817481994628906
+      ],
+      "gap64_mm": [
+        0.5,
+        0.5,
+        0.5,
+        -0.08817155079043282
+      ],
+      "pair32": [
+        0,
+        0,
+        0,
+        6
+      ],
+      "pair64": [
+        0,
+        0,
+        0,
+        6
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        34.57734298706055,
+        32.23564910888672,
+        -71.76512908935547,
+        10.783411979675293,
+        88.44934844970703,
+        -105.16633605957031
+      ],
+      "endpoint": [
+        34.31901931762695,
+        33.24732208251953,
+        -72.12022399902344,
+        10.171412467956543,
+        89.0,
+        -106.89080810546875
+      ],
+      "dense_nsegments": 173,
+      "dense_gap64_min_mm": -0.08817155079043282
+    },
+    {
+      "i": 266,
+      "task": {
+        "tid": 458057184,
+        "id": "s0005_000271",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          44.394,
+          32.402,
+          -145.074,
+          -48.37,
+          -85.543,
+          13.576
+        ],
+        "T": [
+          184.38,
+          -306.445,
+          433.207
+        ],
+        "R": null,
+        "exec_s": 2.52,
+        "travel": 376.205,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 1.2200000286102295,
+      "dls_it": 304,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.5,
+        0.2372283935546875,
+        -0.05915641784667969
+      ],
+      "gap64_mm": [
+        0.5,
+        0.5,
+        0.23719947210205383,
+        -0.05918908805933398
+      ],
+      "pair32": [
+        0,
+        0,
+        5,
+        5
+      ],
+      "pair64": [
+        0,
+        0,
+        5,
+        5
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        20.49032974243164,
+        10.35391902923584,
+        -172.87489318847656,
+        -73.93492126464844,
+        -93.70915985107422,
+        14.725995063781738
+      ],
+      "endpoint": [
+        20.34522819519043,
+        10.012239456176758,
+        -172.91363525390625,
+        -74.4600601196289,
+        -94.0,
+        13.266871452331543
+      ],
+      "dense_nsegments": 146,
+      "dense_gap64_min_mm": -0.05918908805933398
+    },
+    {
+      "i": 325,
+      "task": {
+        "tid": 1206824904,
+        "id": "s0005_001755",
+        "cat": "B",
+        "mode": "pose",
+        "q0": [
+          -68.225,
+          58.481,
+          -31.386,
+          57.2,
+          -31.831,
+          -80.259
+        ],
+        "T": [
+          30.805,
+          353.752,
+          661.419
+        ],
+        "R": [
+          0.5493387160259536,
+          0.13571391775309405,
+          0.8245051289125273,
+          -0.09525700289583322,
+          -0.9701190077749987,
+          0.22314841283987508,
+          0.8301524429129026,
+          -0.2011239500451364,
+          -0.5199962290670785
+        ],
+        "exec_s": 2.34,
+        "travel": 659.756,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 3.0999999046325684,
+      "dls_it": -1,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.3628082275390625,
+        0.004294395446777344,
+        -0.6861848831176758
+      ],
+      "gap64_mm": [
+        0.5,
+        0.3628058862147938,
+        0.004307730814957722,
+        -0.686172816414107
+      ],
+      "pair32": [
+        0,
+        18,
+        5,
+        5
+      ],
+      "pair64": [
+        0,
+        18,
+        5,
+        5
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        -39.524147033691406,
+        54.68175506591797,
+        -18.574026107788086,
+        103.78845977783203,
+        -94.0,
+        -53.657958984375
+      ],
+      "endpoint": [
+        -39.42676544189453,
+        55.778987884521484,
+        -18.88759994506836,
+        103.97470092773438,
+        -94.0,
+        -54.84250259399414
+      ],
+      "dense_nsegments": 119,
+      "dense_gap64_min_mm": -0.686172816414107
+    }
+  ]
+}

docs/experiment_round30_after.json

@@ -0,0 +1,62 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round30_after_rows.json",
+    "out": "docs/experiment_round30_after.json",
+    "trace": "ml/runs/round30_after.pt",
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.57421875,
+    "collision_free_success": 0.57421875,
+    "time_ok_1p3": 0.390625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.4449527263641357,
+    "time_mean_s": 2.3076188564300537,
+    "brakes_mean": 19.25,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.20997506380081177,
+    "err_fail_med": 229.68746948242188,
+    "err_success_med": 0.18899640440940857,
+    "dls_share": 0.625,
+    "time_budget": "teacher×2(≥3 s)"
+  },
+  "simulation_wall_s": 248.83432349999202
+}

docs/experiment_round30_before.json

@@ -0,0 +1,62 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round30_before_rows.json",
+    "out": "docs/experiment_round30_before.json",
+    "trace": "ml/runs/round30_before.pt",
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.57421875,
+    "collision_free_success": 0.56640625,
+    "time_ok_1p3": 0.390625,
+    "collision_rate": 0.02734375,
+    "collision_steps_mean": 0.0625,
+    "travel_ratio_median": 2.4449527263641357,
+    "time_mean_s": 2.2993197441101074,
+    "brakes_mean": 19.23828125,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.20997506380081177,
+    "err_fail_med": 229.68746948242188,
+    "err_success_med": 0.1950554996728897,
+    "dls_share": 0.625,
+    "time_budget": "teacher×2(≥3 s)"
+  },
+  "simulation_wall_s": 247.51454629999353
+}

docs/experiment_round30_certificates.json

@@ -0,0 +1,880 @@
+{
+  "trace": "ml/runs/round30_before.pt",
+  "n": 256,
+  "reported_collision_tasks": 7,
+  "replayed_first_hits32": 7,
+  "replayed_first_hits64": 7,
+  "scope": "Replay reported first hits only; same geometry kernel at float32/float64, not a full safety audit.",
+  "rows": [
+    {
+      "i": 21,
+      "task": {
+        "tid": 1223280991,
+        "id": "s0001_000564",
+        "cat": "A",
+        "mode": "pose",
+        "q0": [
+          178.767,
+          101.917,
+          25.283,
+          31.588,
+          -12.179,
+          -6.464
+        ],
+        "T": [
+          320.779,
+          79.294,
+          350.853
+        ],
+        "R": [
+          -0.5548626629953249,
+          -0.1382477148217311,
+          -0.820374910976869,
+          -0.29956278115972873,
+          -0.8867511636570162,
+          0.35204334093514156,
+          -0.7761375943462505,
+          0.44108949556693955,
+          0.4506112421399604
+        ],
+        "exec_s": 2.86,
+        "travel": 581.023,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 5.659999847412109,
+      "dls_it": -1,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.5,
+        0.35770511627197266,
+        -0.3957242965698242
+      ],
+      "gap64_mm": [
+        0.5,
+        0.5,
+        0.3577231796451059,
+        -0.39572636491366353
+      ],
+      "pair32": [
+        0,
+        0,
+        5,
+        5
+      ],
+      "pair64": [
+        0,
+        0,
+        5,
+        5
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        139.78468322753906,
+        87.256591796875,
+        6.928672790527344,
+        103.78778839111328,
+        -94.0,
+        19.4688720703125
+      ],
+      "endpoint": [
+        139.50450134277344,
+        87.4175796508789,
+        6.71871280670166,
+        104.29376983642578,
+        -94.0,
+        21.218488693237305
+      ],
+      "dense_nsegments": 175,
+      "dense_gap64_min_mm": -0.39572636491366353,
+      "certificate": {
+        "plan_start_it": 282,
+        "step": 0,
+        "checked": true,
+        "endpoint": [
+          139.50450134277344,
+          87.4175796508789,
+          6.71871280670166,
+          104.29376983642578,
+          -95.717529296875,
+          21.218488693237305
+        ],
+        "midpoint": [
+          139.64459228515625,
+          87.33708190917969,
+          6.823692798614502,
+          104.04077911376953,
+          -94.8587646484375,
+          20.34368133544922
+        ],
+        "recorded_gap32_mm": [
+          0.1606760025024414,
+          0.5
+        ],
+        "replayed_gap64_mm": [
+          0.16066912536394184,
+          0.5
+        ],
+        "command_delta": [
+          -0.2801789343357086,
+          0.16098570823669434,
+          -0.20995977520942688,
+          0.5059784054756165,
+          -1.717526912689209,
+          1.7496168613433838
+        ],
+        "clamped_reconstruction_max_abs_deg": 0.0,
+        "certified_vs_executed_max_abs_deg": 1.717529296875
+      }
+    },
+    {
+      "i": 84,
+      "task": {
+        "tid": 1268963905,
+        "id": "s0002_000538",
+        "cat": "B",
+        "mode": "pose",
+        "q0": [
+          -49.198,
+          52.395,
+          108.796,
+          75.534,
+          82.302,
+          -121.823
+        ],
+        "T": [
+          -85.062,
+          275.378,
+          888.929
+        ],
+        "R": [
+          0.5139412462799992,
+          -0.4285630099475821,
+          -0.7431003578769356,
+          -0.5533754886400825,
+          -0.8275489878032067,
+          0.09454227286372051,
+          -0.6554692700230204,
+          0.36262435010704513,
+          -0.6624678231921397
+        ],
+        "exec_s": 3.2600000000000002,
+        "travel": 349.09,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 5.980000019073486,
+      "dls_it": -1,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.030836105346679688,
+        -0.1565418243408203,
+        -0.11790847778320312
+      ],
+      "gap64_mm": [
+        0.5,
+        0.030838235678215753,
+        -0.15655898804830848,
+        -0.11791701257236653
+      ],
+      "pair32": [
+        0,
+        6,
+        6,
+        6
+      ],
+      "pair64": [
+        0,
+        6,
+        6,
+        6
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        -69.87329864501953,
+        36.559112548828125,
+        2.094588279724121,
+        102.22724151611328,
+        89.0,
+        -212.1793212890625
+      ],
+      "endpoint": [
+        -69.79582214355469,
+        37.351837158203125,
+        -0.033895254135131836,
+        102.45509338378906,
+        89.0,
+        -214.75440979003906
+      ],
+      "dense_nsegments": 258,
+      "dense_gap64_min_mm": -0.16882067924099076,
+      "certificate": {
+        "plan_start_it": 298,
+        "step": 0,
+        "checked": true,
+        "endpoint": [
+          -69.79582214355469,
+          37.351837158203125,
+          -0.033895254135131836,
+          102.45509338378906,
+          89.79981231689453,
+          -214.75440979003906
+        ],
+        "midpoint": [
+          -69.83456420898438,
+          36.955474853515625,
+          1.0303465127944946,
+          102.34117126464844,
+          89.39990234375,
+          -213.4668731689453
+        ],
+        "recorded_gap32_mm": [
+          0.5,
+          0.20721912384033203
+        ],
+        "replayed_gap64_mm": [
+          0.5,
+          0.2071966639119509
+        ],
+        "command_delta": [
+          0.07747504115104675,
+          0.7927232980728149,
+          -2.128483533859253,
+          0.22785267233848572,
+          0.7998088598251343,
+          -2.5750951766967773
+        ],
+        "clamped_reconstruction_max_abs_deg": 0.0,
+        "certified_vs_executed_max_abs_deg": 0.7998123168945312
+      }
+    },
+    {
+      "i": 145,
+      "task": {
+        "tid": 593213945,
+        "id": "s0003_000414",
+        "cat": "B",
+        "mode": "pose",
+        "q0": [
+          -49.395,
+          84.783,
+          2.744,
+          14.841,
+          -86.783,
+          81.932
+        ],
+        "T": [
+          5.391,
+          506.723,
+          383.367
+        ],
+        "R": [
+          -0.5269859333114162,
+          -0.7554822816510317,
+          -0.38927156099983334,
+          0.6402732880654818,
+          -0.05172298162073713,
+          -0.7664038424760652,
+          0.5588702377848656,
+          -0.6531242265324508,
+          0.5109724082909718
+        ],
+        "exec_s": 2.96,
+        "travel": 560.55,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 5.599999904632568,
+      "dls_it": -1,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.5,
+        0.3180522918701172,
+        -0.19827747344970703
+      ],
+      "gap64_mm": [
+        0.5,
+        0.5,
+        0.318071359587325,
+        -0.19825289671655533
+      ],
+      "pair32": [
+        0,
+        0,
+        5,
+        5
+      ],
+      "pair64": [
+        0,
+        0,
+        5,
+        5
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        -36.05080032348633,
+        49.852294921875,
+        22.054914474487305,
+        67.3380126953125,
+        -94.0,
+        79.87274932861328
+      ],
+      "endpoint": [
+        -35.60316848754883,
+        48.9223747253418,
+        21.386226654052734,
+        67.87915802001953,
+        -94.0,
+        79.27398681640625
+      ],
+      "dense_nsegments": 93,
+      "dense_gap64_min_mm": -0.19825289671655533,
+      "certificate": {
+        "plan_start_it": 278,
+        "step": 1,
+        "checked": true,
+        "endpoint": [
+          -35.60316848754883,
+          48.9223747253418,
+          21.3862247467041,
+          67.87915802001953,
+          -97.17477416992188,
+          79.27399444580078
+        ],
+        "midpoint": [
+          -35.82698440551758,
+          49.387332916259766,
+          21.720569610595703,
+          67.60858154296875,
+          -96.70620727539062,
+          79.57337188720703
+        ],
+        "recorded_gap32_mm": [
+          0.2614564895629883,
+          0.5
+        ],
+        "replayed_gap64_mm": [
+          0.2614702651611598,
+          0.5
+        ],
+        "command_delta": [
+          0.4476320445537567,
+          -0.929920494556427,
+          -0.6686887741088867,
+          0.5411425828933716,
+          -0.9371238946914673,
+          -0.5987589359283447
+        ],
+        "clamped_reconstruction_max_abs_deg": 0.0,
+        "certified_vs_executed_max_abs_deg": 3.174774169921875
+      }
+    },
+    {
+      "i": 164,
+      "task": {
+        "tid": 1743026973,
+        "id": "s0003_000814",
+        "cat": "C",
+        "mode": "pos",
+        "q0": [
+          -65.982,
+          5.595,
+          -64.669,
+          -72.912,
+          85.732,
+          -86.377
+        ],
+        "T": [
+          -189.354,
+          221.034,
+          302.457
+        ],
+        "R": null,
+        "exec_s": 2.3000000000000003,
+        "travel": 306.824,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 3.0399999618530273,
+      "dls_it": -1,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.5,
+        0.5,
+        -0.12672901153564453
+      ],
+      "gap64_mm": [
+        0.5,
+        0.5,
+        0.5,
+        -0.12673094091899273
+      ],
+      "pair32": [
+        0,
+        0,
+        0,
+        5
+      ],
+      "pair64": [
+        0,
+        0,
+        0,
+        5
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        -33.6672248840332,
+        39.63850021362305,
+        -60.174774169921875,
+        -108.2870864868164,
+        88.27082061767578,
+        -74.56200408935547
+      ],
+      "endpoint": [
+        -32.8936882019043,
+        39.478729248046875,
+        -58.639915466308594,
+        -107.65913391113281,
+        89.0,
+        -74.92655944824219
+      ],
+      "dense_nsegments": 154,
+      "dense_gap64_min_mm": -0.12673094091899273,
+      "certificate": {
+        "plan_start_it": 150,
+        "step": 1,
+        "checked": true,
+        "endpoint": [
+          -32.89369201660156,
+          39.478729248046875,
+          -58.639915466308594,
+          -107.65913391113281,
+          89.33978271484375,
+          -74.92655944824219
+        ],
+        "midpoint": [
+          -33.28045654296875,
+          39.55861282348633,
+          -59.407344818115234,
+          -107.97311401367188,
+          88.8052978515625,
+          -74.7442855834961
+        ],
+        "recorded_gap32_mm": [
+          0.1356372833251953,
+          0.5
+        ],
+        "replayed_gap64_mm": [
+          0.13560383338968585,
+          0.5
+        ],
+        "command_delta": [
+          0.773534893989563,
+          -0.15977251529693604,
+          1.534859299659729,
+          0.6279491186141968,
+          1.068959355354309,
+          -0.3645569384098053
+        ],
+        "clamped_reconstruction_max_abs_deg": 0.0,
+        "certified_vs_executed_max_abs_deg": 0.33978271484375
+      }
+    },
+    {
+      "i": 175,
+      "task": {
+        "tid": 444406422,
+        "id": "s0003_001100",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          -28.854,
+          43.33,
+          -137.874,
+          -6.895,
+          -88.001,
+          -43.745
+        ],
+        "T": [
+          379.105,
+          293.88,
+          653.541
+        ],
+        "R": null,
+        "exec_s": 2.58,
+        "travel": 465.057,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 0.8399999737739563,
+      "dls_it": 144,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.4698514938354492,
+        -0.6612262725830078,
+        -0.9359521865844727
+      ],
+      "gap64_mm": [
+        0.5,
+        0.4698356780126254,
+        -0.6612412101671552,
+        -0.9359581587234
+      ],
+      "pair32": [
+        0,
+        6,
+        6,
+        6
+      ],
+      "pair64": [
+        0,
+        6,
+        6,
+        6
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        -42.48794937133789,
+        11.756612777709961,
+        -124.7530288696289,
+        -46.12489700317383,
+        -94.0,
+        -54.57075500488281
+      ],
+      "endpoint": [
+        -42.73639678955078,
+        10.6148681640625,
+        -124.47537231445312,
+        -48.52891159057617,
+        -94.0,
+        -55.451263427734375
+      ],
+      "dense_nsegments": 241,
+      "dense_gap64_min_mm": -0.9544276230224238,
+      "certificate": {
+        "plan_start_it": 40,
+        "step": 1,
+        "checked": true,
+        "endpoint": [
+          -42.73639678955078,
+          10.6148681640625,
+          -124.47537231445312,
+          -48.52891159057617,
+          -98.31655883789062,
+          -55.45126724243164
+        ],
+        "midpoint": [
+          -42.61217498779297,
+          11.18574047088623,
+          -124.61420440673828,
+          -47.326904296875,
+          -97.59821319580078,
+          -55.011009216308594
+        ],
+        "recorded_gap32_mm": [
+          0.5,
+          0.5
+        ],
+        "replayed_gap64_mm": [
+          0.5,
+          0.5
+        ],
+        "command_delta": [
+          -0.24844767153263092,
+          -1.1417450904846191,
+          0.27765607833862305,
+          -2.4040162563323975,
+          -1.4366930723190308,
+          -0.8805096745491028
+        ],
+        "clamped_reconstruction_max_abs_deg": 0.0,
+        "certified_vs_executed_max_abs_deg": 4.316558837890625
+      }
+    },
+    {
+      "i": 180,
+      "task": {
+        "tid": 737760595,
+        "id": "s0003_001280",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          41.213,
+          49.353,
+          -149.808,
+          -8.981,
+          -88.566,
+          83.45
+        ],
+        "T": [
+          118.145,
+          2.733,
+          266.082
+        ],
+        "R": null,
+        "exec_s": 1.78,
+        "travel": 228.67,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 1.1399999856948853,
+      "dls_it": 102,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.06517887115478516,
+        -0.0172576904296875,
+        0.013133049011230469
+      ],
+      "gap64_mm": [
+        0.5,
+        0.06521150921177998,
+        -0.017227151210720493,
+        0.013145878955246104
+      ],
+      "pair32": [
+        0,
+        5,
+        5,
+        5
+      ],
+      "pair64": [
+        0,
+        5,
+        5,
+        5
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        1.8212896585464478,
+        -9.327469825744629,
+        -150.87704467773438,
+        -23.00735855102539,
+        -94.0,
+        93.73664855957031
+      ],
+      "endpoint": [
+        1.3598560094833374,
+        -9.66404914855957,
+        -150.52992248535156,
+        -23.200042724609375,
+        -94.0,
+        92.69589233398438
+      ],
+      "dense_nsegments": 105,
+      "dense_gap64_min_mm": -0.020246891242656773,
+      "certificate": {
+        "plan_start_it": 56,
+        "step": 0,
+        "checked": true,
+        "endpoint": [
+          1.3598560094833374,
+          -9.66404914855957,
+          -150.52992248535156,
+          -23.200042724609375,
+          -95.86970520019531,
+          92.69589233398438
+        ],
+        "midpoint": [
+          1.5905728340148926,
+          -9.495759010314941,
+          -150.70347595214844,
+          -23.103700637817383,
+          -94.93485260009766,
+          93.21627044677734
+        ],
+        "recorded_gap32_mm": [
+          0.5,
+          0.20821666717529297
+        ],
+        "replayed_gap64_mm": [
+          0.5,
+          0.20822553884000428
+        ],
+        "command_delta": [
+          -0.46143364906311035,
+          -0.33657926321029663,
+          0.3471294939517975,
+          -0.19268345832824707,
+          -1.8697090148925781,
+          -1.0407572984695435
+        ],
+        "clamped_reconstruction_max_abs_deg": 0.0,
+        "certified_vs_executed_max_abs_deg": 1.8697052001953125
+      }
+    },
+    {
+      "i": 215,
+      "task": {
+        "tid": 169730876,
+        "id": "s0004_000544",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          47.051,
+          -1.179,
+          -94.31,
+          -43.128,
+          67.067,
+          -119.867
+        ],
+        "T": [
+          -142.696,
+          -139.159,
+          726.163
+        ],
+        "R": null,
+        "exec_s": 2.16,
+        "travel": 461.684,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "first_hit_s": 3.5399999618530273,
+      "dls_it": -1,
+      "point_order": [
+        "q0",
+        "previous",
+        "midpoint",
+        "endpoint"
+      ],
+      "gap32_mm": [
+        0.5,
+        0.5,
+        0.5,
+        -0.08817481994628906
+      ],
+      "gap64_mm": [
+        0.5,
+        0.5,
+        0.5,
+        -0.08817155079043282
+      ],
+      "pair32": [
+        0,
+        0,
+        0,
+        6
+      ],
+      "pair64": [
+        0,
+        0,
+        0,
+        6
+      ],
+      "boundary_joints": [
+        5
+      ],
+      "previous": [
+        34.57734298706055,
+        32.23564910888672,
+        -71.76512908935547,
+        10.783411979675293,
+        88.44934844970703,
+        -105.16633605957031
+      ],
+      "endpoint": [
+        34.31901931762695,
+        33.24732208251953,
+        -72.12022399902344,
+        10.171412467956543,
+        89.0,
+        -106.89080810546875
+      ],
+      "dense_nsegments": 173,
+      "dense_gap64_min_mm": -0.08817155079043282,
+      "certificate": {
+        "plan_start_it": 176,
+        "step": 0,
+        "checked": true,
+        "endpoint": [
+          34.31901931762695,
+          33.24732208251953,
+          -72.12022399902344,
+          10.171412467956543,
+          89.79435729980469,
+          -106.89080810546875
+        ],
+        "midpoint": [
+          34.44818115234375,
+          32.741485595703125,
+          -71.94267272949219,
+          10.477412223815918,
+          89.12185668945312,
+          -106.02857208251953
+        ],
+        "recorded_gap32_mm": [
+          0.2809743881225586,
+          0.5
+        ],
+        "replayed_gap64_mm": [
+          0.28098007564065064,
+          0.5
+        ],
+        "command_delta": [
+          -0.2583238482475281,
+          1.011674404144287,
+          -0.3550943434238434,
+          -0.6119998693466187,
+          1.3450126647949219,
+          -1.7244750261306763
+        ],
+        "clamped_reconstruction_max_abs_deg": 0.0,
+        "certified_vs_executed_max_abs_deg": 0.7943572998046875
+      }
+    }
+  ]
+}

docs/experiment_round30_comparison.json

@@ -0,0 +1,65 @@
+{
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "same_config_except_output_paths": true,
+  "instrumentation_vs_round29_first_batch_max_abs_deg": 0.0,
+  "reached_only_after": [],
+  "reached_only_before": [],
+  "collision_tasks_before": [
+    21,
+    84,
+    145,
+    164,
+    175,
+    180,
+    215
+  ],
+  "collision_tasks_after": [],
+  "statistics": {
+    "before": {
+      "active_certified_steps": 34069,
+      "certificate_endpoint_mismatch_gt_0p001deg": 5298,
+      "certificate_endpoint_max_abs_deg": 7.587158203125,
+      "certificate_midpoint_max_abs_deg": 5.78558349609375,
+      "uncertified_moving_steps": 0,
+      "executed_speed_ratio_max": 1.660003662109375,
+      "executed_speed_limit_violating_tasks": 256,
+      "executed_speed_max_deg_s": [
+        99.6002197265625,
+        99.60002899169922,
+        149.40032958984375,
+        149.400146484375,
+        199.20005798339844,
+        298.80029296875
+      ],
+      "executed_travel_ratio_median_success": 2.4250845909118652,
+      "commanded_travel_ratio_median_success": 2.4449527263641357
+    },
+    "after": {
+      "active_certified_steps": 34107,
+      "certificate_endpoint_mismatch_gt_0p001deg": 0,
+      "certificate_endpoint_max_abs_deg": 0.0,
+      "certificate_midpoint_max_abs_deg": 0.0,
+      "uncertified_moving_steps": 0,
+      "executed_speed_ratio_max": 1.660003662109375,
+      "executed_speed_limit_violating_tasks": 256,
+      "executed_speed_max_deg_s": [
+        99.6002197265625,
+        99.60002899169922,
+        149.40032958984375,
+        149.400146484375,
+        199.20005798339844,
+        298.80029296875
+      ],
+      "executed_travel_ratio_median_success": 2.4250845909118652,
+      "commanded_travel_ratio_median_success": 2.4449527263641357
+    }
+  },
+  "cpu_prediction_microbenchmark": {
+    "scope": "B=1, commit=2, 100 warmups then 1000 serial calls, CPU one thread; not full planning latency",
+    "before_p50_ms": 0.0388,
+    "before_p99_ms": 0.069102,
+    "after_p50_ms": 0.0518,
+    "after_p99_ms": 0.093933
+  }
+}

docs/experiment_round30_dense_after.json

@@ -0,0 +1,485 @@
+{
+  "trace": "ml/runs/round30_after.pt",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "indices": [
+    0,
+    21,
+    84,
+    128,
+    145,
+    164,
+    175,
+    180,
+    197,
+    215,
+    224,
+    255
+  ],
+  "max_step_deg": 0.05,
+  "dtype": "float64",
+  "device": "cuda",
+  "scope": "Executed trajectory through each task deadline or arrival, including initial state; finite sampling, not continuous certification.",
+  "collision_tasks": 1,
+  "rows": [
+    {
+      "i": 0,
+      "task": {
+        "tid": 919549409,
+        "id": "s0001_000001",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          -129.897,
+          46.68,
+          43.494,
+          -97.461,
+          -85.102,
+          -139.71
+        ],
+        "T": [
+          -116.172,
+          268.408,
+          179.282
+        ],
+        "R": null,
+        "exec_s": 3.48,
+        "travel": 345.324,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 174,
+      "sampled_points": 6211,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.02673023799951224,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 21,
+      "task": {
+        "tid": 1223280991,
+        "id": "s0001_000564",
+        "cat": "A",
+        "mode": "pose",
+        "q0": [
+          178.767,
+          101.917,
+          25.283,
+          31.588,
+          -12.179,
+          -6.464
+        ],
+        "T": [
+          320.779,
+          79.294,
+          350.853
+        ],
+        "R": [
+          -0.5548626629953249,
+          -0.1382477148217311,
+          -0.820374910976869,
+          -0.29956278115972873,
+          -0.8867511636570162,
+          0.35204334093514156,
+          -0.7761375943462505,
+          0.44108949556693955,
+          0.4506112421399604
+        ],
+        "exec_s": 2.86,
+        "travel": 581.023,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 286,
+      "sampled_points": 6059,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.009057896844284485,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 84,
+      "task": {
+        "tid": 1268963905,
+        "id": "s0002_000538",
+        "cat": "B",
+        "mode": "pose",
+        "q0": [
+          -49.198,
+          52.395,
+          108.796,
+          75.534,
+          82.302,
+          -121.823
+        ],
+        "T": [
+          -85.062,
+          275.378,
+          888.929
+        ],
+        "R": [
+          0.5139412462799992,
+          -0.4285630099475821,
+          -0.7431003578769356,
+          -0.5533754886400825,
+          -0.8275489878032067,
+          0.09454227286372051,
+          -0.6554692700230204,
+          0.36262435010704513,
+          -0.6624678231921397
+        ],
+        "exec_s": 3.2600000000000002,
+        "travel": 349.09,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 326,
+      "sampled_points": 6251,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.006206321604070197,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 128,
+      "task": {
+        "tid": 408952743,
+        "id": "s0002_001588",
+        "cat": "A",
+        "mode": "pos",
+        "q0": [
+          -153.539,
+          100.454,
+          108.434,
+          3.379,
+          70.254,
+          -179.848
+        ],
+        "T": [
+          -324.009,
+          70.778,
+          919.477
+        ],
+        "R": null,
+        "exec_s": 0.8200000000000001,
+        "travel": 121.056,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 63,
+      "sampled_points": 2403,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.5,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 145,
+      "task": {
+        "tid": 593213945,
+        "id": "s0003_000414",
+        "cat": "B",
+        "mode": "pose",
+        "q0": [
+          -49.395,
+          84.783,
+          2.744,
+          14.841,
+          -86.783,
+          81.932
+        ],
+        "T": [
+          5.391,
+          506.723,
+          383.367
+        ],
+        "R": [
+          -0.5269859333114162,
+          -0.7554822816510317,
+          -0.38927156099983334,
+          0.6402732880654818,
+          -0.05172298162073713,
+          -0.7664038424760652,
+          0.5588702377848656,
+          -0.6531242265324508,
+          0.5109724082909718
+        ],
+        "exec_s": 2.96,
+        "travel": 560.55,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 296,
+      "sampled_points": 4509,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.002145835308116517,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 164,
+      "task": {
+        "tid": 1743026973,
+        "id": "s0003_000814",
+        "cat": "C",
+        "mode": "pos",
+        "q0": [
+          -65.982,
+          5.595,
+          -64.669,
+          -72.912,
+          85.732,
+          -86.377
+        ],
+        "T": [
+          -189.354,
+          221.034,
+          302.457
+        ],
+        "R": null,
+        "exec_s": 2.3000000000000003,
+        "travel": 306.824,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 230,
+      "sampled_points": 3985,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.008857213584009926,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 175,
+      "task": {
+        "tid": 444406422,
+        "id": "s0003_001100",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          -28.854,
+          43.33,
+          -137.874,
+          -6.895,
+          -88.001,
+          -43.745
+        ],
+        "T": [
+          379.105,
+          293.88,
+          653.541
+        ],
+        "R": null,
+        "exec_s": 2.58,
+        "travel": 465.057,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 143,
+      "sampled_points": 5547,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.18812628629569694,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 180,
+      "task": {
+        "tid": 737760595,
+        "id": "s0003_001280",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          41.213,
+          49.353,
+          -149.808,
+          -8.981,
+          -88.566,
+          83.45
+        ],
+        "T": [
+          118.145,
+          2.733,
+          266.082
+        ],
+        "R": null,
+        "exec_s": 1.78,
+        "travel": 228.67,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 95,
+      "sampled_points": 3153,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": -0.1492959159660998,
+      "collision_segments": 1,
+      "first_collision_s": 1.0915789473684212
+    },
+    {
+      "i": 197,
+      "task": {
+        "tid": 1433345185,
+        "id": "s0004_000020",
+        "cat": "B",
+        "mode": "pose",
+        "q0": [
+          20.112,
+          -20.013,
+          -177.85,
+          32.746,
+          -70.888,
+          -30.111
+        ],
+        "T": [
+          -288.364,
+          63.606,
+          672.54
+        ],
+        "R": [
+          -0.5335375732231524,
+          -0.7509294450860695,
+          0.38915630081225366,
+          -0.8443344656004893,
+          0.4997538915602813,
+          -0.19324947107169785,
+          -0.049365657681080755,
+          -0.4316839311036288,
+          -0.9006730902323173
+        ],
+        "exec_s": 2.14,
+        "travel": 634.022,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 214,
+      "sampled_points": 4641,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.0010684324191032601,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 215,
+      "task": {
+        "tid": 169730876,
+        "id": "s0004_000544",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          47.051,
+          -1.179,
+          -94.31,
+          -43.128,
+          67.067,
+          -119.867
+        ],
+        "T": [
+          -142.696,
+          -139.159,
+          726.163
+        ],
+        "R": null,
+        "exec_s": 2.16,
+        "travel": 461.684,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 216,
+      "sampled_points": 6605,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.2112382958686112,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 224,
+      "task": {
+        "tid": 1884410379,
+        "id": "s0004_000721",
+        "cat": "A",
+        "mode": "pose",
+        "q0": [
+          -127.447,
+          7.845,
+          -124.228,
+          -20.968,
+          -21.51,
+          -35.505
+        ],
+        "T": [
+          -233.158,
+          -59.78,
+          648.987
+        ],
+        "R": [
+          0.3964618999442966,
+          -0.020495209950317686,
+          0.9178223729358808,
+          -0.8179090164279511,
+          0.4461640898622928,
+          0.3632663564978371,
+          -0.41694460392049726,
+          -0.8947164641865291,
+          0.1601238457417738
+        ],
+        "exec_s": 1.48,
+        "travel": 356.988,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 128,
+      "sampled_points": 4707,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.0041216155493710716,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 255,
+      "task": {
+        "tid": 712677274,
+        "id": "s0004_001686",
+        "cat": "C",
+        "mode": "pos",
+        "q0": [
+          -65.548,
+          71.865,
+          27.462,
+          -75.048,
+          60.325,
+          151.872
+        ],
+        "T": [
+          -53.154,
+          50.204,
+          397.489
+        ],
+        "R": null,
+        "exec_s": 1.84,
+        "travel": 238.182,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 184,
+      "sampled_points": 2495,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.004616787795093558,
+      "collision_segments": 0,
+      "first_collision_s": null
+    }
+  ],
+  "audit_wall_s": 15.056756100006169
+}

docs/experiment_round30_dense_before.json

@@ -0,0 +1,485 @@
+{
+  "trace": "ml/runs/round30_before.pt",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "indices": [
+    0,
+    21,
+    84,
+    128,
+    145,
+    164,
+    175,
+    180,
+    197,
+    215,
+    224,
+    255
+  ],
+  "max_step_deg": 0.05,
+  "dtype": "float64",
+  "device": "cuda",
+  "scope": "Executed trajectory through each task deadline or arrival, including initial state; finite sampling, not continuous certification.",
+  "collision_tasks": 7,
+  "rows": [
+    {
+      "i": 0,
+      "task": {
+        "tid": 919549409,
+        "id": "s0001_000001",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          -129.897,
+          46.68,
+          43.494,
+          -97.461,
+          -85.102,
+          -139.71
+        ],
+        "T": [
+          -116.172,
+          268.408,
+          179.282
+        ],
+        "R": null,
+        "exec_s": 3.48,
+        "travel": 345.324,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 174,
+      "sampled_points": 6211,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.02673023799951224,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 21,
+      "task": {
+        "tid": 1223280991,
+        "id": "s0001_000564",
+        "cat": "A",
+        "mode": "pose",
+        "q0": [
+          178.767,
+          101.917,
+          25.283,
+          31.588,
+          -12.179,
+          -6.464
+        ],
+        "T": [
+          320.779,
+          79.294,
+          350.853
+        ],
+        "R": [
+          -0.5548626629953249,
+          -0.1382477148217311,
+          -0.820374910976869,
+          -0.29956278115972873,
+          -0.8867511636570162,
+          0.35204334093514156,
+          -0.7761375943462505,
+          0.44108949556693955,
+          0.4506112421399604
+        ],
+        "exec_s": 2.86,
+        "travel": 581.023,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 286,
+      "sampled_points": 6099,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": -0.39572636491366353,
+      "collision_segments": 2,
+      "first_collision_s": 5.655
+    },
+    {
+      "i": 84,
+      "task": {
+        "tid": 1268963905,
+        "id": "s0002_000538",
+        "cat": "B",
+        "mode": "pose",
+        "q0": [
+          -49.198,
+          52.395,
+          108.796,
+          75.534,
+          82.302,
+          -121.823
+        ],
+        "T": [
+          -85.062,
+          275.378,
+          888.929
+        ],
+        "R": [
+          0.5139412462799992,
+          -0.4285630099475821,
+          -0.7431003578769356,
+          -0.5533754886400825,
+          -0.8275489878032067,
+          0.09454227286372051,
+          -0.6554692700230204,
+          0.36262435010704513,
+          -0.6624678231921397
+        ],
+        "exec_s": 3.2600000000000002,
+        "travel": 349.09,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 326,
+      "sampled_points": 6037,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": -0.16878710314368206,
+      "collision_segments": 3,
+      "first_collision_s": 5.961153846153847
+    },
+    {
+      "i": 128,
+      "task": {
+        "tid": 408952743,
+        "id": "s0002_001588",
+        "cat": "A",
+        "mode": "pos",
+        "q0": [
+          -153.539,
+          100.454,
+          108.434,
+          3.379,
+          70.254,
+          -179.848
+        ],
+        "T": [
+          -324.009,
+          70.778,
+          919.477
+        ],
+        "R": null,
+        "exec_s": 0.8200000000000001,
+        "travel": 121.056,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 63,
+      "sampled_points": 2403,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.5,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 145,
+      "task": {
+        "tid": 593213945,
+        "id": "s0003_000414",
+        "cat": "B",
+        "mode": "pose",
+        "q0": [
+          -49.395,
+          84.783,
+          2.744,
+          14.841,
+          -86.783,
+          81.932
+        ],
+        "T": [
+          5.391,
+          506.723,
+          383.367
+        ],
+        "R": [
+          -0.5269859333114162,
+          -0.7554822816510317,
+          -0.38927156099983334,
+          0.6402732880654818,
+          -0.05172298162073713,
+          -0.7664038424760652,
+          0.5588702377848656,
+          -0.6531242265324508,
+          0.5109724082909718
+        ],
+        "exec_s": 2.96,
+        "travel": 560.55,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 296,
+      "sampled_points": 4489,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": -0.19825289671655533,
+      "collision_segments": 5,
+      "first_collision_s": 5.597
+    },
+    {
+      "i": 164,
+      "task": {
+        "tid": 1743026973,
+        "id": "s0003_000814",
+        "cat": "C",
+        "mode": "pos",
+        "q0": [
+          -65.982,
+          5.595,
+          -64.669,
+          -72.912,
+          85.732,
+          -86.377
+        ],
+        "T": [
+          -189.354,
+          221.034,
+          302.457
+        ],
+        "R": null,
+        "exec_s": 2.3000000000000003,
+        "travel": 306.824,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 230,
+      "sampled_points": 3865,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": -0.12673094091895543,
+      "collision_segments": 4,
+      "first_collision_s": 3.0387500000000003
+    },
+    {
+      "i": 175,
+      "task": {
+        "tid": 444406422,
+        "id": "s0003_001100",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          -28.854,
+          43.33,
+          -137.874,
+          -6.895,
+          -88.001,
+          -43.745
+        ],
+        "T": [
+          379.105,
+          293.88,
+          653.541
+        ],
+        "R": null,
+        "exec_s": 2.58,
+        "travel": 465.057,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 145,
+      "sampled_points": 5769,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": -0.954427674326439,
+      "collision_segments": 2,
+      "first_collision_s": 0.8236
+    },
+    {
+      "i": 180,
+      "task": {
+        "tid": 737760595,
+        "id": "s0003_001280",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          41.213,
+          49.353,
+          -149.808,
+          -8.981,
+          -88.566,
+          83.45
+        ],
+        "T": [
+          118.145,
+          2.733,
+          266.082
+        ],
+        "R": null,
+        "exec_s": 1.78,
+        "travel": 228.67,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 103,
+      "sampled_points": 3385,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": -0.1492959159660998,
+      "collision_segments": 2,
+      "first_collision_s": 1.0915789473684212
+    },
+    {
+      "i": 197,
+      "task": {
+        "tid": 1433345185,
+        "id": "s0004_000020",
+        "cat": "B",
+        "mode": "pose",
+        "q0": [
+          20.112,
+          -20.013,
+          -177.85,
+          32.746,
+          -70.888,
+          -30.111
+        ],
+        "T": [
+          -288.364,
+          63.606,
+          672.54
+        ],
+        "R": [
+          -0.5335375732231524,
+          -0.7509294450860695,
+          0.38915630081225366,
+          -0.8443344656004893,
+          0.4997538915602813,
+          -0.19324947107169785,
+          -0.049365657681080755,
+          -0.4316839311036288,
+          -0.9006730902323173
+        ],
+        "exec_s": 2.14,
+        "travel": 634.022,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 214,
+      "sampled_points": 4317,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.0010684324191032601,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 215,
+      "task": {
+        "tid": 169730876,
+        "id": "s0004_000544",
+        "cat": "B",
+        "mode": "pos",
+        "q0": [
+          47.051,
+          -1.179,
+          -94.31,
+          -43.128,
+          67.067,
+          -119.867
+        ],
+        "T": [
+          -142.696,
+          -139.159,
+          726.163
+        ],
+        "R": null,
+        "exec_s": 2.16,
+        "travel": 461.684,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 216,
+      "sampled_points": 6341,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": -0.08817155079042394,
+      "collision_segments": 3,
+      "first_collision_s": 3.538888888888889
+    },
+    {
+      "i": 224,
+      "task": {
+        "tid": 1884410379,
+        "id": "s0004_000721",
+        "cat": "A",
+        "mode": "pose",
+        "q0": [
+          -127.447,
+          7.845,
+          -124.228,
+          -20.968,
+          -21.51,
+          -35.505
+        ],
+        "T": [
+          -233.158,
+          -59.78,
+          648.987
+        ],
+        "R": [
+          0.3964618999442966,
+          -0.020495209950317686,
+          0.9178223729358808,
+          -0.8179090164279511,
+          0.4461640898622928,
+          0.3632663564978371,
+          -0.41694460392049726,
+          -0.8947164641865291,
+          0.1601238457417738
+        ],
+        "exec_s": 1.48,
+        "travel": 356.988,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 128,
+      "sampled_points": 4707,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.0041216155493710716,
+      "collision_segments": 0,
+      "first_collision_s": null
+    },
+    {
+      "i": 255,
+      "task": {
+        "tid": 712677274,
+        "id": "s0004_001686",
+        "cat": "C",
+        "mode": "pos",
+        "q0": [
+          -65.548,
+          71.865,
+          27.462,
+          -75.048,
+          60.325,
+          151.872
+        ],
+        "T": [
+          -53.154,
+          50.204,
+          397.489
+        ],
+        "R": null,
+        "exec_s": 1.84,
+        "travel": 238.182,
+        "reached": true,
+        "err": 0,
+        "split": "val"
+      },
+      "steps": 184,
+      "sampled_points": 2495,
+      "initial_gap64_mm": 0.5,
+      "min_gap64_mm": 0.004616787795093558,
+      "collision_segments": 0,
+      "first_collision_s": null
+    }
+  ],
+  "audit_wall_s": 14.869853199998033
+}

docs/experiment_round30_missed_segment.json

@@ -0,0 +1,71 @@
+{
+  "task_index": 180,
+  "task": {
+    "tid": 737760595,
+    "id": "s0003_001280",
+    "cat": "B",
+    "mode": "pos",
+    "q0": [
+      41.213,
+      49.353,
+      -149.808,
+      -8.981,
+      -88.566,
+      83.45
+    ],
+    "T": [
+      118.145,
+      2.733,
+      266.082
+    ],
+    "R": null,
+    "exec_s": 1.78,
+    "travel": 228.67,
+    "reached": true,
+    "err": 0,
+    "split": "val"
+  },
+  "segment_start_it": 54,
+  "start_s": 1.08,
+  "end_s": 1.1,
+  "original_discrete_col": 0,
+  "certified": true,
+  "recorded_endpoint_midpoint_gap32_mm": [
+    0.0651998519897461,
+    0.11720466613769531
+  ],
+  "previous_midpoint_endpoint_gap64_mm": [
+    0.5,
+    0.11723191087430074,
+    0.06521150921177998
+  ],
+  "previous": [
+    3.0118656158447266,
+    -7.4926042556762695,
+    -151.498291015625,
+    -22.7498836517334,
+    -94.0,
+    93.50477600097656
+  ],
+  "endpoint": [
+    1.8212896585464478,
+    -9.327469825744629,
+    -150.87704467773438,
+    -23.00735855102539,
+    -94.0,
+    93.73664855957031
+  ],
+  "dense_max_step_deg": 0.01,
+  "dense_points": 185,
+  "min_gap64_mm": -0.14927288174659914,
+  "min_at_fraction": 0.7663043478260869,
+  "min_at_q": [
+    2.09952208334985,
+    -8.898669719696045,
+    -151.0222272458284,
+    -22.947187786516935,
+    -94.0,
+    93.68246095076852
+  ],
+  "scope": "One executed segment; finite sampling demonstrates a collision between certified checkpoints."
+}

docs/experiment_round31_cert_cost_model.json

@@ -0,0 +1,658 @@
+{
+  "segments": 48,
+  "kin_fp": "2ebcbf17",
+  "root_gap_cap_mm": 20.0,
+  "formula": "L = ceil(log2(bound_pair/(ga+gb)))\uff0c\u67e5\u8be2\u6570 = 2^L + 1\uff1bbound_pair \u53d6\u6700\u7d27\u7ec4\u5bf9\uff0c\u4e0d\u7528\u5168\u7ec4\u5bf9\u6700\u5927\u503c",
+  "levels_needed_by_status": {
+    "safe": [
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      0,
+      1,
+      1,
+      1,
+      2,
+      2,
+      2,
+      2,
+      3,
+      3,
+      3,
+      3,
+      3,
+      4,
+      4,
+      4,
+      4,
+      4
+    ],
+    "unresolved": [
+      5,
+      5,
+      5,
+      5,
+      5,
+      6,
+      6,
+      6,
+      6,
+      6
+    ],
+    "collision": [
+      4
+    ]
+  },
+  "queries_needed_safe": [
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    3,
+    3,
+    3,
+    5,
+    5,
+    5,
+    5,
+    9,
+    9,
+    9,
+    9,
+    9,
+    17,
+    17,
+    17,
+    17,
+    17
+  ],
+  "model_vs_measured_max_excess": 20,
+  "unresolved_needed_queries": [
+    33,
+    33,
+    33,
+    33,
+    33,
+    65,
+    65,
+    65,
+    65,
+    65
+  ],
+  "collision_dense_check": [
+    {
+      "i": 180,
+      "start_it": 54,
+      "status": "collision",
+      "tight_group_pair": 5,
+      "bound_pair_mm": 29.09,
+      "endpoint_gap_sum_mm": 1.98854,
+      "levels_needed": 4,
+      "queries_needed": 17,
+      "queries_used_batch32": 5
+    }
+  ],
+  "note": "\u6a21\u578b\u662f\u201c\u754c\u53ea\u968f\u4e8c\u5206\u51cf\u534a\u201d\u4e0b\u7684\u4e0b\u754c\uff1b\u5b9e\u6d4b\u67e5\u8be2\u53ef\u80fd\u66f4\u5c0f\uff08\u591a\u5c42\u91cc\u6709\u7684\u7ec4\u5bf9\u5148\u6ee1\u8db3\uff09\u6216\u66f4\u5927\uff08\u78b0\u649e\u6bb5\u63d0\u524d\u7ec8\u6b62\uff09\u3002",
+  "rows": [
+    {
+      "i": 0,
+      "start_it": 0,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 2.382,
+      "endpoint_gap_sum_mm": 17.77168,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 0,
+      "start_it": 8,
+      "status": "unresolved",
+      "tight_group_pair": 13,
+      "bound_pair_mm": 8.2833,
+      "endpoint_gap_sum_mm": 0.25216,
+      "levels_needed": 6,
+      "queries_needed": 65,
+      "queries_used_batch32": 17
+    },
+    {
+      "i": 0,
+      "start_it": 91,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 2.459,
+      "endpoint_gap_sum_mm": 14.45459,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 0,
+      "start_it": 173,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 6.8776,
+      "endpoint_gap_sum_mm": 15.48695,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 21,
+      "start_it": 0,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 4.7095,
+      "endpoint_gap_sum_mm": 12.81844,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 21,
+      "start_it": 100,
+      "status": "unresolved",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 1.3371,
+      "endpoint_gap_sum_mm": 0.03365,
+      "levels_needed": 6,
+      "queries_needed": 65,
+      "queries_used_batch32": 17
+    },
+    {
+      "i": 21,
+      "start_it": 127,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 1.5104,
+      "endpoint_gap_sum_mm": 0.68046,
+      "levels_needed": 2,
+      "queries_needed": 5,
+      "queries_used_batch32": 4
+    },
+    {
+      "i": 21,
+      "start_it": 282,
+      "status": "safe",
+      "tight_group_pair": 5,
+      "bound_pair_mm": 3.5692,
+      "endpoint_gap_sum_mm": 1.72327,
+      "levels_needed": 2,
+      "queries_needed": 5,
+      "queries_used_batch32": 4
+    },
+    {
+      "i": 84,
+      "start_it": 0,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 11.7042,
+      "endpoint_gap_sum_mm": 13.79742,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 84,
+      "start_it": 170,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 4.9891,
+      "endpoint_gap_sum_mm": 0.3999,
+      "levels_needed": 4,
+      "queries_needed": 17,
+      "queries_used_batch32": 18
+    },
+    {
+      "i": 84,
+      "start_it": 234,
+      "status": "unresolved",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 7.5939,
+      "endpoint_gap_sum_mm": 0.28698,
+      "levels_needed": 5,
+      "queries_needed": 33,
+      "queries_used_batch32": 31
+    },
+    {
+      "i": 84,
+      "start_it": 325,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 3.5738,
+      "endpoint_gap_sum_mm": 10.95173,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 128,
+      "start_it": 0,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 4.4648,
+      "endpoint_gap_sum_mm": 16.52712,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 128,
+      "start_it": 31,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 2.1155,
+      "endpoint_gap_sum_mm": 14.24909,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 128,
+      "start_it": 62,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 0.3619,
+      "endpoint_gap_sum_mm": 12.62201,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 145,
+      "start_it": 0,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 1.5691,
+      "endpoint_gap_sum_mm": 11.60111,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 145,
+      "start_it": 32,
+      "status": "safe",
+      "tight_group_pair": 4,
+      "bound_pair_mm": 4.6053,
+      "endpoint_gap_sum_mm": 0.67129,
+      "levels_needed": 3,
+      "queries_needed": 9,
+      "queries_used_batch32": 28
+    },
+    {
+      "i": 145,
+      "start_it": 138,
+      "status": "safe",
+      "tight_group_pair": 4,
+      "bound_pair_mm": 9.6062,
+      "endpoint_gap_sum_mm": 1.11765,
+      "levels_needed": 4,
+      "queries_needed": 17,
+      "queries_used_batch32": 15
+    },
+    {
+      "i": 145,
+      "start_it": 294,
+      "status": "safe",
+      "tight_group_pair": 5,
+      "bound_pair_mm": 8.2956,
+      "endpoint_gap_sum_mm": 0.61574,
+      "levels_needed": 4,
+      "queries_needed": 17,
+      "queries_used_batch32": 17
+    },
+    {
+      "i": 164,
+      "start_it": 0,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 12.0298,
+      "endpoint_gap_sum_mm": 11.50625,
+      "levels_needed": 1,
+      "queries_needed": 3,
+      "queries_used_batch32": 3
+    },
+    {
+      "i": 164,
+      "start_it": 84,
+      "status": "safe",
+      "tight_group_pair": 5,
+      "bound_pair_mm": 18.0143,
+      "endpoint_gap_sum_mm": 1.7857,
+      "levels_needed": 4,
+      "queries_needed": 17,
+      "queries_used_batch32": 18
+    },
+    {
+      "i": 164,
+      "start_it": 182,
+      "status": "unresolved",
+      "tight_group_pair": 5,
+      "bound_pair_mm": 14.1522,
+      "endpoint_gap_sum_mm": 0.38976,
+      "levels_needed": 6,
+      "queries_needed": 65,
+      "queries_used_batch32": 17
+    },
+    {
+      "i": 164,
+      "start_it": 227,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 5.6457,
+      "endpoint_gap_sum_mm": 1.13586,
+      "levels_needed": 3,
+      "queries_needed": 9,
+      "queries_used_batch32": 10
+    },
+    {
+      "i": 175,
+      "start_it": 0,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 6.5456,
+      "endpoint_gap_sum_mm": 15.18067,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 175,
+      "start_it": 46,
+      "status": "unresolved",
+      "tight_group_pair": 6,
+      "bound_pair_mm": 30.8668,
+      "endpoint_gap_sum_mm": 1.26529,
+      "levels_needed": 5,
+      "queries_needed": 33,
+      "queries_used_batch32": 32
+    },
+    {
+      "i": 175,
+      "start_it": 75,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 17.8885,
+      "endpoint_gap_sum_mm": 10.74989,
+      "levels_needed": 1,
+      "queries_needed": 3,
+      "queries_used_batch32": 3
+    },
+    {
+      "i": 175,
+      "start_it": 142,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 0.8185,
+      "endpoint_gap_sum_mm": 10.77208,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 180,
+      "start_it": 0,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 3.4654,
+      "endpoint_gap_sum_mm": 14.04707,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 180,
+      "start_it": 14,
+      "status": "unresolved",
+      "tight_group_pair": 4,
+      "bound_pair_mm": 17.6609,
+      "endpoint_gap_sum_mm": 0.46562,
+      "levels_needed": 6,
+      "queries_needed": 65,
+      "queries_used_batch32": 17
+    },
+    {
+      "i": 180,
+      "start_it": 52,
+      "status": "safe",
+      "tight_group_pair": 5,
+      "bound_pair_mm": 27.7231,
+      "endpoint_gap_sum_mm": 3.97272,
+      "levels_needed": 3,
+      "queries_needed": 9,
+      "queries_used_batch32": 15
+    },
+    {
+      "i": 180,
+      "start_it": 54,
+      "status": "collision",
+      "tight_group_pair": 5,
+      "bound_pair_mm": 29.09,
+      "endpoint_gap_sum_mm": 1.98854,
+      "levels_needed": 4,
+      "queries_needed": 17,
+      "queries_used_batch32": 5
+    },
+    {
+      "i": 180,
+      "start_it": 94,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 1.1032,
+      "endpoint_gap_sum_mm": 12.22495,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 197,
+      "start_it": 0,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 1.0839,
+      "endpoint_gap_sum_mm": 6.87225,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 197,
+      "start_it": 80,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 3.3325,
+      "endpoint_gap_sum_mm": 0.50053,
+      "levels_needed": 3,
+      "queries_needed": 9,
+      "queries_used_batch32": 29
+    },
+    {
+      "i": 197,
+      "start_it": 82,
+      "status": "unresolved",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 5.7899,
+      "endpoint_gap_sum_mm": 0.16354,
+      "levels_needed": 6,
+      "queries_needed": 65,
+      "queries_used_batch32": 17
+    },
+    {
+      "i": 197,
+      "start_it": 211,
+      "status": "safe",
+      "tight_group_pair": 5,
+      "bound_pair_mm": 18.6183,
+      "endpoint_gap_sum_mm": 5.55945,
+      "levels_needed": 2,
+      "queries_needed": 5,
+      "queries_used_batch32": 11
+    },
+    {
+      "i": 215,
+      "start_it": 0,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 2.4256,
+      "endpoint_gap_sum_mm": 12.47554,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 215,
+      "start_it": 104,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 4.7208,
+      "endpoint_gap_sum_mm": 14.00267,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 215,
+      "start_it": 178,
+      "status": "unresolved",
+      "tight_group_pair": 6,
+      "bound_pair_mm": 25.7963,
+      "endpoint_gap_sum_mm": 1.07014,
+      "levels_needed": 5,
+      "queries_needed": 33,
+      "queries_used_batch32": 30
+    },
+    {
+      "i": 215,
+      "start_it": 215,
+      "status": "safe",
+      "tight_group_pair": 6,
+      "bound_pair_mm": 9.2152,
+      "endpoint_gap_sum_mm": 8.57509,
+      "levels_needed": 1,
+      "queries_needed": 3,
+      "queries_used_batch32": 3
+    },
+    {
+      "i": 224,
+      "start_it": 0,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 6.492,
+      "endpoint_gap_sum_mm": 10.02154,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 224,
+      "start_it": 52,
+      "status": "unresolved",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 3.5989,
+      "endpoint_gap_sum_mm": 0.12354,
+      "levels_needed": 5,
+      "queries_needed": 33,
+      "queries_used_batch32": 32
+    },
+    {
+      "i": 224,
+      "start_it": 61,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 2.7788,
+      "endpoint_gap_sum_mm": 0.78961,
+      "levels_needed": 2,
+      "queries_needed": 5,
+      "queries_used_batch32": 5
+    },
+    {
+      "i": 224,
+      "start_it": 127,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 1.4884,
+      "endpoint_gap_sum_mm": 7.29024,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 255,
+      "start_it": 0,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 1.5332,
+      "endpoint_gap_sum_mm": 11.184,
+      "levels_needed": 0,
+      "queries_needed": 2,
+      "queries_used_batch32": 2
+    },
+    {
+      "i": 255,
+      "start_it": 25,
+      "status": "unresolved",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 2.9831,
+      "endpoint_gap_sum_mm": 0.14423,
+      "levels_needed": 5,
+      "queries_needed": 33,
+      "queries_used_batch32": 29
+    },
+    {
+      "i": 255,
+      "start_it": 50,
+      "status": "safe",
+      "tight_group_pair": 4,
+      "bound_pair_mm": 7.2365,
+      "endpoint_gap_sum_mm": 0.61112,
+      "levels_needed": 4,
+      "queries_needed": 17,
+      "queries_used_batch32": 25
+    },
+    {
+      "i": 255,
+      "start_it": 181,
+      "status": "safe",
+      "tight_group_pair": 18,
+      "bound_pair_mm": 4.2961,
+      "endpoint_gap_sum_mm": 0.74979,
+      "levels_needed": 3,
+      "queries_needed": 9,
+      "queries_used_batch32": 12
+    }
+  ]
+}

docs/experiment_round31_distance.json

@@ -0,0 +1,24 @@
+{
+  "baseline_commit": "95da491",
+  "poses": "first 64 existing parity poses",
+  "n_poses": 64,
+  "kin_fp": "2ebcbf17",
+  "dtype": "float64",
+  "degenerate_capsules": [
+    686
+  ],
+  "pairs_with_second_degenerate": 670,
+  "evaluated_pairs": 42880,
+  "changed_pairs_over_1e_8": 20661,
+  "old_minus_new_min_mm": 0.0,
+  "old_minus_new_max_mm": 86.00309853133399,
+  "new_reverse_max_abs_mm": 0.0,
+  "worst": {
+    "pose": 29,
+    "first_capsule": 430,
+    "second_capsule": 686,
+    "old_distance_mm": 493.354753665718,
+    "new_distance_mm": 407.35165513438403
+  },
+  "scope": "Raw capsule distances; does not imply these pairs win the group minimum."
+}

docs/experiment_round31_kernel_replay.json

@@ -0,0 +1,16 @@
+{
+  "baseline_commit": "95da491",
+  "sample_source": "docs/experiment_round31_certificates.json",
+  "segments": 48,
+  "dense_points": 7586,
+  "max_step_deg": 0.01,
+  "cap_mm": 0.5,
+  "dtype": "float64",
+  "device": "cuda",
+  "kin_fp": "2ebcbf17",
+  "max_group_gap_difference_mm": 0.0,
+  "group_gaps_changed_over_1e_8": 0,
+  "minimum_gaps_changed_over_1e_8": 0,
+  "collision_bool_mismatches": 0,
+  "wall_s": 4.18259759999637
+}

docs/experiment_round31_parity.json

@@ -0,0 +1,10 @@
+{
+ "n": 2000,
+ "fk_max_abs_mm": 4.547473508864641e-13,
+ "orient_max_abs": 3.3306690738754696e-16,
+ "gap_max_abs_mm": 2.5224267119483557e-13,
+ "collide_bool_mismatch": 0,
+ "feat_max_abs": 5.960463655974024e-08,
+ "grad_rel_err_max": 1.8474076070450906e-06,
+ "PASS": true
+}

docs/experiment_round32_bound_falsification.json

@@ -0,0 +1,22 @@
+{
+  "scope": "全部 48 段 × 15 组对 = 720 项检验；观察值按 cap=2.0 mm 钳制，故可检验命题取 min(预测下限, cap)。违例=实测最小间隙低于该下限。",
+  "checks": 720,
+  "violations": [],
+  "worst_margin_mm": 0.0,
+  "min_slack_mm": 1e-06,
+  "dense_cap_mm": 2.0,
+  "grid_points_per_segment": 121,
+  "supersedes": "docs/experiment_round32_uniform_validation.json 里 bound_checks=61 的部分覆盖版本",
+  "kin_fp": "2ebcbf17",
+  "wall_s": 1.66,
+  "nontrivial": {
+    "definition": "只有预测下限 < 观察 cap 才是有信息量的检验；其余组对被钳制恒等式自动满足",
+    "checks": 61,
+    "min_slack_mm": 0.035347,
+    "p10_slack_mm": 0.544274,
+    "violations": 0,
+    "pred_mm_median_all": 15.8639,
+    "pred_mm_max_all": 19.9582
+  },
+  "note": "720 项总检验 0 违例，其中有信息量的是 61 项；预测下限的中位数 15.9 mm 远大于观察 cap 2 mm，所以大部分组对无法在本网格上证伪"
+}

docs/experiment_round32_uniform_latency.json

@@ -0,0 +1,190 @@
+{
+  "device": "cuda float64",
+  "torch": "2.10.0+cu130",
+  "kin_fp": "2ebcbf17",
+  "roots_per_call": 2,
+  "pairs": 24,
+  "warm": 3,
+  "reps": 7,
+  "segment_sample": "docs/experiment_round31_certificates.json 的 48 段两两成对",
+  "design": "每对根轮流按正序/逆序测完全部配置（3次预热+7次计时）；uniform_auto 的宽相位=最大子段界+2eps，uniform_cap20 固定 20 mm，bisect 为第31轮逐层二分口径",
+  "results": [
+    {
+      "kind": "uniform_auto",
+      "n_or_budget": 8,
+      "points": 18,
+      "cap_used_mm": 5.9191,
+      "calls": 168,
+      "p50_ms": 9.684,
+      "p90_ms": 13.027,
+      "p99_ms": 15.157,
+      "max_ms": 17.537,
+      "over_20ms": 0,
+      "roots": 48,
+      "safe": 26,
+      "collision": 1,
+      "rejected": 21
+    },
+    {
+      "kind": "uniform_auto",
+      "n_or_budget": 16,
+      "points": 34,
+      "cap_used_mm": 2.9597,
+      "calls": 168,
+      "p50_ms": 14.137,
+      "p90_ms": 18.454,
+      "p99_ms": 20.866,
+      "max_ms": 21.632,
+      "over_20ms": 5,
+      "roots": 48,
+      "safe": 30,
+      "collision": 1,
+      "rejected": 17
+    },
+    {
+      "kind": "uniform_auto",
+      "n_or_budget": 32,
+      "points": 66,
+      "cap_used_mm": 1.4799,
+      "calls": 168,
+      "p50_ms": 20.209,
+      "p90_ms": 27.88,
+      "p99_ms": 32.268,
+      "max_ms": 35.563,
+      "over_20ms": 87,
+      "roots": 48,
+      "safe": 34,
+      "collision": 1,
+      "rejected": 13
+    },
+    {
+      "kind": "uniform_auto",
+      "n_or_budget": 64,
+      "points": 130,
+      "cap_used_mm": 0.7401,
+      "calls": 168,
+      "p50_ms": 31.429,
+      "p90_ms": 48.751,
+      "p99_ms": 57.697,
+      "max_ms": 62.084,
+      "over_20ms": 165,
+      "roots": 48,
+      "safe": 36,
+      "collision": 1,
+      "rejected": 11
+    },
+    {
+      "kind": "uniform_auto",
+      "n_or_budget": 128,
+      "points": 258,
+      "cap_used_mm": 0.3701,
+      "calls": 168,
+      "p50_ms": 59.761,
+      "p90_ms": 89.063,
+      "p99_ms": 109.086,
+      "max_ms": 117.345,
+      "over_20ms": 168,
+      "roots": 48,
+      "safe": 38,
+      "collision": 1,
+      "rejected": 9
+    },
+    {
+      "kind": "uniform_cap20",
+      "n_or_budget": 8,
+      "points": 18,
+      "cap_used_mm": 20.0,
+      "calls": 168,
+      "p50_ms": 13.129,
+      "p90_ms": 16.44,
+      "p99_ms": 20.963,
+      "max_ms": 24.67,
+      "over_20ms": 3,
+      "roots": 48,
+      "safe": 26,
+      "collision": 1,
+      "rejected": 21
+    },
+    {
+      "kind": "uniform_cap20",
+      "n_or_budget": 32,
+      "points": 66,
+      "cap_used_mm": 20.0,
+      "calls": 168,
+      "p50_ms": 28.415,
+      "p90_ms": 39.579,
+      "p99_ms": 53.023,
+      "max_ms": 59.983,
+      "over_20ms": 153,
+      "roots": 48,
+      "safe": 34,
+      "collision": 1,
+      "rejected": 13
+    },
+    {
+      "kind": "uniform_cap20",
+      "n_or_budget": 64,
+      "points": 130,
+      "cap_used_mm": 20.0,
+      "calls": 168,
+      "p50_ms": 50.85,
+      "p90_ms": 75.512,
+      "p99_ms": 88.355,
+      "max_ms": 91.492,
+      "over_20ms": 168,
+      "roots": 48,
+      "safe": 36,
+      "collision": 1,
+      "rejected": 11
+    },
+    {
+      "kind": "uniform_cap20",
+      "n_or_budget": 128,
+      "points": 258,
+      "cap_used_mm": 20.0,
+      "calls": 168,
+      "p50_ms": 98.46,
+      "p90_ms": 139.982,
+      "p99_ms": 167.042,
+      "max_ms": 177.743,
+      "over_20ms": 168,
+      "roots": 48,
+      "safe": 38,
+      "collision": 1,
+      "rejected": 9
+    },
+    {
+      "kind": "bisect",
+      "n_or_budget": 32,
+      "points": null,
+      "cap_used_mm": 20.0,
+      "calls": 168,
+      "p50_ms": 54.885,
+      "p90_ms": 82.151,
+      "p99_ms": 117.205,
+      "max_ms": 118.716,
+      "over_20ms": 133,
+      "roots": 48,
+      "safe": 37,
+      "collision": 1,
+      "rejected": 10
+    },
+    {
+      "kind": "bisect",
+      "n_or_budget": 65,
+      "points": null,
+      "cap_used_mm": 20.0,
+      "calls": 168,
+      "p50_ms": 71.825,
+      "p90_ms": 108.559,
+      "p99_ms": 122.533,
+      "max_ms": 127.319,
+      "over_20ms": 133,
+      "roots": 48,
+      "safe": 41,
+      "collision": 1,
+      "rejected": 6
+    }
+  ],
+  "wall_s": 108.63
+}

docs/experiment_round32_uniform_validation.json

@@ -0,0 +1,62 @@
+{
+  "segments": 48,
+  "kin_fp": "2ebcbf17",
+  "dense_cap_mm": 2.0,
+  "eps_mm": 0.0001,
+  "bound_falsification": "每条组对：预测下限 (ga+gb-bound)/2 应 ≤ 网格上该组对的最小实测间隙",
+  "per_segment": [
+    {
+      "n": 8,
+      "safe": 26,
+      "collision": 1,
+      "unproven": 21,
+      "outside_box": 0,
+      "false_accept_vs_dense": 0,
+      "cap_required_mm": 5.919107314806348,
+      "known_missed_segment_collision": true
+    },
+    {
+      "n": 16,
+      "safe": 30,
+      "collision": 1,
+      "unproven": 17,
+      "outside_box": 0,
+      "false_accept_vs_dense": 0,
+      "cap_required_mm": 2.959653657403174,
+      "known_missed_segment_collision": true
+    },
+    {
+      "n": 32,
+      "safe": 34,
+      "collision": 1,
+      "unproven": 13,
+      "outside_box": 0,
+      "false_accept_vs_dense": 0,
+      "cap_required_mm": 1.479926828701587,
+      "known_missed_segment_collision": true
+    },
+    {
+      "n": 64,
+      "safe": 36,
+      "collision": 1,
+      "unproven": 11,
+      "outside_box": 0,
+      "false_accept_vs_dense": 0,
+      "cap_required_mm": 0.7400634143507935,
+      "known_missed_segment_collision": true
+    },
+    {
+      "n": 128,
+      "safe": 38,
+      "collision": 1,
+      "unproven": 9,
+      "outside_box": 0,
+      "false_accept_vs_dense": 0,
+      "cap_required_mm": 0.37013170717539673,
+      "known_missed_segment_collision": true
+    }
+  ],
+  "bound_checks": 61,
+  "bound_violations": [],
+  "worst_margin_mm": 0.0
+}

docs/experiment_round33_base.json

@@ -0,0 +1,70 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round33_base_rows.json",
+    "out": "docs/experiment_round33_base.json",
+    "trace": "ml/runs/round33_base.pt",
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 0,
+    "gap_margin": 0.0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.57421875,
+    "collision_free_success": 0.57421875,
+    "time_ok_1p3": 0.390625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.4449527263641357,
+    "time_mean_s": 2.3076188564300537,
+    "brakes_mean": 19.25,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.20997506380081177,
+    "err_fail_med": 229.68746948242188,
+    "err_success_med": 0.18899640440940857,
+    "dls_share": 0.625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.6600004434585571,
+    "speed_ratio_p95": 1.6600035429000854,
+    "speed_ratio_max": 1.660003662109375,
+    "over_vmax_rate": 1.0,
+    "accel_ratio_p95": 1.5835107564926147,
+    "accel_ratio_max": 1.6600004434585571
+  },
+  "simulation_wall_s": 296.2702505999914
+}

docs/experiment_round33_both.json

@@ -0,0 +1,70 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round33_both_rows.json",
+    "out": "docs/experiment_round33_both.json",
+    "trace": "ml/runs/round33_both.pt",
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.5,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.51953125,
+    "collision_free_success": 0.51953125,
+    "time_ok_1p3": 0.26171875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.5444939136505127,
+    "time_mean_s": 2.4729323387145996,
+    "brakes_mean": 19.97265625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2868920564651489,
+    "err_fail_med": 202.12567138671875,
+    "err_success_med": 0.22085000574588776,
+    "dls_share": 0.62109375,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118
+  },
+  "simulation_wall_s": 277.54961390000244
+}

docs/experiment_round33_cap.json

@@ -0,0 +1,70 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round33_cap_rows.json",
+    "out": "docs/experiment_round33_cap.json",
+    "trace": "ml/runs/round33_cap.pt",
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5234375,
+    "collision_free_success": 0.5234375,
+    "time_ok_1p3": 0.265625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.615795850753784,
+    "time_mean_s": 2.486567258834839,
+    "brakes_mean": 19.5546875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.27920001745224,
+    "err_fail_med": 204.71360778808594,
+    "err_success_med": 0.21751396358013153,
+    "dls_share": 0.61328125,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118
+  },
+  "simulation_wall_s": 325.18856850000157
+}

docs/experiment_round33_gpu_cycle.json

@@ -0,0 +1,202 @@
+{
+  "device": "NVIDIA GeForce RTX 3090",
+  "torch": "2.10.0+cu130",
+  "act_scale": 1.0,
+  "cfg": {},
+  "B": 1,
+  "commit_n": 2,
+  "fp64_note": "3090 fp64 \u541e\u5410\u7ea6 fp32 \u7684 1/54\uff1b\u8ba4\u8bc1\u76ee\u524d\u8981\u6c42 fp64",
+  "inference": {
+    "sample_K1_steps1": {
+      "p50": 7.453,
+      "p95": 9.912,
+      "p99": 10.766,
+      "max": 11.109,
+      "reps": 120
+    },
+    "sample_K8_steps1": {
+      "p50": 8.208,
+      "p95": 10.685,
+      "p99": 11.22,
+      "max": 15.224,
+      "reps": 120
+    },
+    "sample_K1_steps2": {
+      "p50": 13.149,
+      "p95": 17.621,
+      "p99": 18.438,
+      "max": 19.296,
+      "reps": 120
+    },
+    "sample_K8_steps2": {
+      "p50": 14.167,
+      "p95": 19.241,
+      "p99": 24.77,
+      "max": 25.703,
+      "reps": 120
+    },
+    "sample_K1_steps4": {
+      "p50": 19.193,
+      "p95": 22.338,
+      "p99": 23.666,
+      "max": 25.787,
+      "reps": 120
+    },
+    "sample_K8_steps4": {
+      "p50": 21.509,
+      "p95": 24.732,
+      "p99": 35.655,
+      "max": 41.514,
+      "reps": 120
+    },
+    "sample_K1_steps8": {
+      "p50": 40.658,
+      "p95": 55.162,
+      "p99": 60.394,
+      "max": 60.398,
+      "reps": 120
+    },
+    "sample_K8_steps8": {
+      "p50": 40.416,
+      "p95": 47.997,
+      "p99": 62.49,
+      "max": 71.976,
+      "reps": 120
+    },
+    "sample_K1_steps16": {
+      "p50": 72.429,
+      "p95": 90.188,
+      "p99": 107.536,
+      "max": 112.674,
+      "reps": 120
+    },
+    "sample_K8_steps16": {
+      "p50": 79.893,
+      "p95": 99.933,
+      "p99": 105.746,
+      "max": 106.078,
+      "reps": 120
+    }
+  },
+  "encode_once": {
+    "p50": 2.265,
+    "p95": 4.001,
+    "p99": 6.086,
+    "max": 6.136,
+    "reps": 120
+  },
+  "tiny_op_ms": 0.028,
+  "sync_only_ms": 0.005,
+  "components": {
+    "features_f32": {
+      "p50": 12.443,
+      "p95": 14.227,
+      "p99": 19.678,
+      "max": 20.531,
+      "reps": 120
+    },
+    "endpoint_score_K8_f32": {
+      "p50": 4.278,
+      "p95": 5.841,
+      "p99": 6.322,
+      "max": 8.883,
+      "reps": 120
+    },
+    "dls_fb_1step_f64": {
+      "p50": 7.191,
+      "p95": 12.26,
+      "p99": 14.642,
+      "max": 15.384,
+      "reps": 120
+    },
+    "dls_fb_1step_f32": {
+      "p50": 7.619,
+      "p95": 10.142,
+      "p99": 11.597,
+      "max": 12.022,
+      "reps": 120
+    },
+    "discrete_cert_4pts_f32": {
+      "p50": 8.214,
+      "p95": 13.365,
+      "p99": 14.235,
+      "max": 14.991,
+      "reps": 120
+    },
+    "cert_n8_f64_2roots": {
+      "p50": 11.265,
+      "p95": 14.64,
+      "p99": 16.788,
+      "max": 17.761,
+      "reps": 120
+    },
+    "cert_n16_f64_2roots": {
+      "p50": 14.574,
+      "p95": 16.731,
+      "p99": 18.459,
+      "max": 22.346,
+      "reps": 120
+    },
+    "cert_n32_f64_2roots": {
+      "p50": 21.528,
+      "p95": 24.746,
+      "p99": 28.037,
+      "max": 28.066,
+      "reps": 120
+    }
+  },
+  "cycle_K8_steps16_discrete": {
+    "p50": 116.474,
+    "p95": 134.474,
+    "p99": 142.593,
+    "max": 152.768,
+    "reps": 120
+  },
+  "cycle_K8_steps4_discrete": {
+    "p50": 55.919,
+    "p95": 65.767,
+    "p99": 74.809,
+    "max": 78.059,
+    "reps": 120
+  },
+  "cycle_K1_steps4_discrete": {
+    "p50": 51.087,
+    "p95": 58.614,
+    "p99": 62.326,
+    "max": 63.919,
+    "reps": 120
+  },
+  "cycle_K1_steps1_discrete": {
+    "p50": 34.493,
+    "p95": 42.078,
+    "p99": 43.906,
+    "max": 50.268,
+    "reps": 120
+  },
+  "cycle_K8_steps16_cert_n8": {
+    "p50": 132.027,
+    "p95": 150.463,
+    "p99": 160.83,
+    "max": 192.586,
+    "reps": 120
+  },
+  "cycle_K8_steps4_cert_n8": {
+    "p50": 64.812,
+    "p95": 79.192,
+    "p99": 89.629,
+    "max": 95.837,
+    "reps": 120
+  },
+  "cycle_K1_steps4_cert_n16": {
+    "p50": 61.967,
+    "p95": 73.219,
+    "p99": 79.053,
+    "max": 83.18,
+    "reps": 120
+  },
+  "f32_vs_f64": {
+    "points": 3,
+    "max_abs_err_mm": 0.0,
+    "sign_flips": 0
+  }
+}

docs/experiment_round33_gpu_cycle_steps1.json

@@ -0,0 +1,143 @@
+{
+  "device": "NVIDIA GeForce RTX 3090",
+  "B": 1,
+  "steps_used_by_eval": 1,
+  "note": "eval_cl \u4ece\u4e0d\u4f20 steps\uff0crollout \u9ed8\u8ba4 steps=1\uff08MeanFlow \u4e00\u6b65\u751f\u6210\uff09\uff1b\u6240\u4ee5\u771f\u5b9e\u5468\u671f\u6309 steps=1 \u6d4b",
+  "parts": {
+    "features_total": {
+      "p50": 12.589,
+      "p95": 18.334,
+      "p99": 20.217,
+      "max": 23.649,
+      "reps": 120
+    },
+    "features_gm_groups_only": {
+      "p50": 6.817,
+      "p95": 8.771,
+      "p99": 10.254,
+      "max": 17.169,
+      "reps": 120
+    },
+    "features_gaps40_all_groups": {
+      "p50": 7.245,
+      "p95": 10.501,
+      "p99": 13.794,
+      "max": 14.508,
+      "reps": 120
+    },
+    "features_fk_only": {
+      "p50": 3.708,
+      "p95": 5.024,
+      "p99": 5.991,
+      "max": 6.301,
+      "reps": 120
+    },
+    "sample_k_K8_steps1": {
+      "p50": 7.799,
+      "p95": 11.655,
+      "p99": 13.342,
+      "max": 16.212,
+      "reps": 120
+    },
+    "sample_k_K1_steps1": {
+      "p50": 7.205,
+      "p95": 8.935,
+      "p99": 9.615,
+      "max": 13.92,
+      "reps": 120
+    },
+    "sample_k_K4_steps1": {
+      "p50": 7.834,
+      "p95": 10.354,
+      "p99": 14.744,
+      "max": 15.241,
+      "reps": 120
+    },
+    "dls_fb_f64": {
+      "p50": 6.729,
+      "p95": 9.297,
+      "p99": 11.186,
+      "max": 12.064,
+      "reps": 120
+    },
+    "discrete_cert_4pts": {
+      "p50": 8.603,
+      "p95": 11.181,
+      "p99": 12.464,
+      "max": 15.799,
+      "reps": 120
+    },
+    "cert_uniform_n8_2roots_f64": {
+      "p50": 11.479,
+      "p95": 19.221,
+      "p99": 20.942,
+      "max": 21.183,
+      "reps": 120
+    },
+    "cert_uniform_n16_2roots_f64": {
+      "p50": 16.138,
+      "p95": 21.972,
+      "p99": 23.898,
+      "max": 25.793,
+      "reps": 120
+    }
+  },
+  "cycles": {
+    "cycle_K8_full_discrete": {
+      "p50": 40.676,
+      "p95": 51.742,
+      "p99": 59.782,
+      "max": 62.074,
+      "reps": 120
+    },
+    "cycle_K4_full_discrete": {
+      "p50": 43.997,
+      "p95": 56.145,
+      "p99": 59.905,
+      "max": 65.882,
+      "reps": 120
+    },
+    "cycle_K1_full_discrete": {
+      "p50": 37.87,
+      "p95": 45.754,
+      "p99": 49.394,
+      "max": 53.139,
+      "reps": 120
+    },
+    "cycle_K8_noscore_discrete": {
+      "p50": 37.427,
+      "p95": 49.602,
+      "p99": 60.777,
+      "max": 70.139,
+      "reps": 120
+    },
+    "cycle_K8_full_cert_n8": {
+      "p50": 43.984,
+      "p95": 65.795,
+      "p99": 75.53,
+      "max": 95.452,
+      "reps": 120
+    },
+    "cycle_K8_full_cert_n16": {
+      "p50": 51.054,
+      "p95": 74.862,
+      "p99": 95.768,
+      "max": 106.48,
+      "reps": 120
+    },
+    "cycle_K8_fused_estimate": {
+      "p50": 24.054,
+      "p95": 26.595,
+      "p99": 30.17,
+      "max": 33.284,
+      "reps": 120
+    },
+    "cycle_K8_fused_cert_n8": {
+      "p50": 34.274,
+      "p95": 46.329,
+      "p99": 72.737,
+      "max": 79.022,
+      "reps": 120
+    }
+  }
+}

docs/experiment_round33_margin.json

@@ -0,0 +1,70 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round33_margin_rows.json",
+    "out": "docs/experiment_round33_margin.json",
+    "trace": "ml/runs/round33_margin.pt",
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 0,
+    "gap_margin": 0.5,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5625,
+    "collision_free_success": 0.5625,
+    "time_ok_1p3": 0.37890625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.448458433151245,
+    "time_mean_s": 2.2555553913116455,
+    "brakes_mean": 20.046875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.18239997327327728,
+    "err_fail_med": 225.54278564453125,
+    "err_success_med": 0.18544802069664001,
+    "dls_share": 0.62890625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.6600004434585571,
+    "speed_ratio_p95": 1.660003662109375,
+    "speed_ratio_max": 1.660003662109375,
+    "over_vmax_rate": 1.0,
+    "accel_ratio_p95": 1.5889430046081543,
+    "accel_ratio_max": 1.6600035429000854
+  },
+  "simulation_wall_s": 298.24517190000915
+}

docs/experiment_round34_cert_batch_check.json

@@ -0,0 +1,71 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round34_cert_batch_check.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 0,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.57421875,
+    "collision_free_success": 0.57421875,
+    "time_ok_1p3": 0.390625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.4449527263641357,
+    "time_mean_s": 2.3076188564300537,
+    "brakes_mean": 19.25,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.20997506380081177,
+    "err_fail_med": 229.68746948242188,
+    "err_success_med": 0.18899640440940857,
+    "dls_share": 0.625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.6600004434585571,
+    "speed_ratio_p95": 1.6600035429000854,
+    "speed_ratio_max": 1.660003662109375,
+    "over_vmax_rate": 1.0,
+    "accel_ratio_p95": 1.5835107564926147,
+    "accel_ratio_max": 1.6600004434585571
+  },
+  "simulation_wall_s": 306.9490996999957
+}

docs/experiment_round34_f32_cert.json

@@ -0,0 +1,41 @@
+{
+  "kin_fp": "2ebcbf17",
+  "segments": 48,
+  "n_subsegments": 8,
+  "wide_cap_mm": 2.0,
+  "gap_abs_err_mm": {
+    "p50": 0.0,
+    "p99": 3.661949063912929e-05,
+    "p999": 7.4360332600075e-05,
+    "max": 0.00013247109886016517
+  },
+  "signed_err_mm": {
+    "min": -0.00010937313134817828,
+    "max": 0.00013247109886016517,
+    "overestimate_count": 1,
+    "overestimate_gt_1e-3": 0
+  },
+  "sign_flips": 0,
+  "accepted_f64": 24,
+  "accepted_f32": 24,
+  "unsafe_direction_f32_only": 0,
+  "conservative_f64_only": 0,
+  "per_group_pair_err_p99": [
+    0.0,
+    7.6e-05,
+    0.0001,
+    7e-05,
+    7.1e-05,
+    0.0,
+    0.0,
+    0.0,
+    0.000132,
+    6.7e-05,
+    5.1e-05,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+  ],
+  "note": "wide_cap 2.0 mm\uff1a\u8ba4\u8bc1\u5b9e\u9645\u7528 bound+2eps \u7684\u81ea\u9002\u5e94 cap\uff0c\u8fd9\u91cc\u53d6\u56fa\u5b9a 2 mm \u4ee5\u8986\u76d6\u88ab\u68c0\u9a8c\u7684\u6bb5\uff1b\u82e5 max \u8bef\u5dee\u8fdc\u5c0f\u4e8e 1e-4 \u7684 eps\uff0c\u5224\u636e\u5728 f32 \u4e0b\u4ecd\u6210\u7acb\u9700\u518d\u786e\u8ba4 f32 \u7684\u95f4\u9699\u672c\u8eab\u8bef\u5dee\u7b26\u53f7\u65b9\u5411\u3002"
+}

docs/experiment_round34_fused_cycle.json

@@ -0,0 +1,629 @@
+{
+  "device": "NVIDIA GeForce RTX 3090",
+  "B": 1,
+  "K": 8,
+  "steps": 1,
+  "reps": 90,
+  "design": "\u540c\u4e00\u6279 4 \u4e2a\u53d8\u4f53\u9010\u8f6e\u4ea4\u66ff\uff08\u5947\u6570\u8f6e\u53cd\u5e8f\uff09\uff0c\u6d88\u9664 GPU \u65f6\u949f\u6f02\u79fb\uff1b\u539f\u59cb\u8017\u65f6\u5168\u90e8\u843d\u76d8",
+  "equivalence": {
+    "features_reuse_bitwise_equal": true,
+    "gm_groups_vs_full_table_max_abs_mm": 0.0,
+    "cert_matrix_bitwise_equal_at_same_cap": true,
+    "shared_xf_cert_bitwise_equal_at_cap1": true,
+    "max_abs_shared_diff_mm": 0.0,
+    "max_abs_cert_diff_mm": 0.0,
+    "decision_equal_by_threshold": {
+      "thr_0.0": true,
+      "thr_0.5": true,
+      "thr_1.0": true
+    },
+    "per_point_min_cur_mm": [
+      8.833421,
+      8.849185,
+      8.859405,
+      8.83829
+    ],
+    "per_point_min_fused_mm": [
+      8.833421,
+      8.849185,
+      8.859405,
+      8.83829
+    ]
+  },
+  "results": {
+    "current_discrete": {
+      "p50": 32.932,
+      "p95": 42.589,
+      "p99": 55.568,
+      "max": 55.568,
+      "hz_p50": 30.4,
+      "raw_ms": [
+        30.458,
+        30.624,
+        30.789,
+        31.17,
+        31.229,
+        31.283,
+        31.371,
+        31.394,
+        31.454,
+        31.473,
+        31.564,
+        31.58,
+        31.719,
+        31.772,
+        31.78,
+        31.785,
+        31.859,
+        31.875,
+        31.906,
+        31.973,
+        31.979,
+        32.002,
+        32.017,
+        32.063,
+        32.099,
+        32.113,
+        32.123,
+        32.127,
+        32.13,
+        32.137,
+        32.191,
+        32.286,
+        32.308,
+        32.324,
+        32.33,
+        32.388,
+        32.48,
+        32.508,
+        32.564,
+        32.63,
+        32.64,
+        32.702,
+        32.78,
+        32.822,
+        32.834,
+        32.932,
+        32.964,
+        32.975,
+        32.996,
+        33.071,
+        33.163,
+        33.191,
+        33.192,
+        33.402,
+        33.477,
+        33.509,
+        33.583,
+        33.956,
+        34.014,
+        34.04,
+        34.047,
+        34.047,
+        34.154,
+        34.302,
+        34.595,
+        34.658,
+        34.943,
+        35.296,
+        35.522,
+        35.67,
+        35.757,
+        35.828,
+        35.859,
+        36.459,
+        36.503,
+        36.562,
+        36.956,
+        37.047,
+        37.067,
+        37.142,
+        37.442,
+        38.632,
+        39.285,
+        41.269,
+        41.932,
+        42.589,
+        43.199,
+        43.329,
+        54.407,
+        55.568
+      ]
+    },
+    "fused_discrete": {
+      "p50": 24.515,
+      "p95": 30.02,
+      "p99": 42.085,
+      "max": 42.085,
+      "hz_p50": 40.8,
+      "raw_ms": [
+        22.146,
+        22.266,
+        22.409,
+        22.653,
+        22.66,
+        22.695,
+        22.76,
+        22.827,
+        22.848,
+        23.001,
+        23.033,
+        23.087,
+        23.249,
+        23.36,
+        23.488,
+        23.573,
+        23.607,
+        23.625,
+        23.649,
+        23.651,
+        23.652,
+        23.652,
+        23.656,
+        23.676,
+        23.716,
+        23.737,
+        23.745,
+        23.795,
+        23.871,
+        23.883,
+        23.92,
+        23.922,
+        23.945,
+        23.959,
+        24.046,
+        24.13,
+        24.184,
+        24.282,
+        24.291,
+        24.313,
+        24.318,
+        24.319,
+        24.32,
+        24.34,
+        24.399,
+        24.515,
+        24.532,
+        24.541,
+        24.552,
+        24.592,
+        24.618,
+        24.731,
+        24.741,
+        24.745,
+        24.766,
+        24.807,
+        24.813,
+        24.898,
+        25.066,
+        25.088,
+        25.137,
+        25.18,
+        25.223,
+        25.376,
+        25.474,
+        25.603,
+        25.68,
+        25.744,
+        25.82,
+        25.936,
+        25.95,
+        25.953,
+        26.002,
+        26.174,
+        26.312,
+        26.88,
+        27.036,
+        27.209,
+        27.329,
+        27.424,
+        27.715,
+        28.796,
+        28.824,
+        29.427,
+        29.662,
+        30.02,
+        31.515,
+        31.839,
+        36.18,
+        42.085
+      ]
+    },
+    "shared_xf_discrete": {
+      "p50": 26.864,
+      "p95": 34.544,
+      "p99": 67.243,
+      "max": 67.243,
+      "hz_p50": 37.2,
+      "raw_ms": [
+        24.433,
+        24.441,
+        24.585,
+        24.777,
+        24.862,
+        24.935,
+        25.054,
+        25.08,
+        25.105,
+        25.193,
+        25.313,
+        25.44,
+        25.466,
+        25.483,
+        25.516,
+        25.554,
+        25.577,
+        25.689,
+        25.702,
+        25.708,
+        25.854,
+        25.896,
+        25.898,
+        26.011,
+        26.097,
+        26.175,
+        26.255,
+        26.289,
+        26.297,
+        26.306,
+        26.328,
+        26.336,
+        26.346,
+        26.375,
+        26.397,
+        26.456,
+        26.502,
+        26.523,
+        26.563,
+        26.613,
+        26.634,
+        26.658,
+        26.708,
+        26.752,
+        26.798,
+        26.864,
+        26.869,
+        26.874,
+        27.315,
+        27.341,
+        27.668,
+        27.67,
+        27.732,
+        27.774,
+        27.812,
+        27.84,
+        27.882,
+        27.921,
+        28.24,
+        28.302,
+        28.421,
+        28.694,
+        28.722,
+        28.739,
+        28.817,
+        28.914,
+        28.944,
+        28.998,
+        29.12,
+        29.122,
+        29.259,
+        29.292,
+        29.334,
+        29.658,
+        29.847,
+        30.181,
+        30.206,
+        30.292,
+        30.492,
+        30.782,
+        31.168,
+        31.222,
+        31.607,
+        31.614,
+        32.187,
+        34.544,
+        35.274,
+        35.565,
+        35.915,
+        67.243
+      ]
+    },
+    "current_cert_n8": {
+      "p50": 42.624,
+      "p95": 49.434,
+      "p99": 57.771,
+      "max": 57.771,
+      "hz_p50": 23.5,
+      "raw_ms": [
+        38.546,
+        38.666,
+        39.152,
+        39.265,
+        39.344,
+        39.441,
+        39.584,
+        39.626,
+        39.738,
+        39.841,
+        40.228,
+        40.306,
+        40.514,
+        40.805,
+        40.838,
+        40.875,
+        40.884,
+        40.907,
+        40.952,
+        40.971,
+        41.185,
+        41.193,
+        41.193,
+        41.277,
+        41.279,
+        41.422,
+        41.448,
+        41.463,
+        41.465,
+        41.473,
+        41.504,
+        41.551,
+        41.684,
+        41.733,
+        41.841,
+        41.947,
+        41.967,
+        42.011,
+        42.077,
+        42.241,
+        42.249,
+        42.269,
+        42.301,
+        42.314,
+        42.388,
+        42.624,
+        42.661,
+        42.767,
+        42.773,
+        42.837,
+        42.865,
+        42.87,
+        42.913,
+        43.343,
+        43.35,
+        43.356,
+        43.376,
+        43.501,
+        43.695,
+        43.708,
+        43.842,
+        43.923,
+        43.972,
+        44.243,
+        44.476,
+        44.536,
+        44.648,
+        44.745,
+        44.894,
+        45.034,
+        45.144,
+        45.236,
+        45.697,
+        45.859,
+        45.944,
+        46.039,
+        46.126,
+        46.467,
+        46.575,
+        46.887,
+        46.983,
+        47.234,
+        47.621,
+        47.685,
+        48.547,
+        49.434,
+        49.683,
+        50.646,
+        54.754,
+        57.771
+      ]
+    },
+    "fused_cert_n8": {
+      "p50": 24.826,
+      "p95": 31.721,
+      "p99": 49.398,
+      "max": 49.398,
+      "hz_p50": 40.3,
+      "raw_ms": [
+        22.127,
+        22.553,
+        22.598,
+        22.624,
+        22.789,
+        22.798,
+        22.878,
+        22.983,
+        23.002,
+        23.134,
+        23.192,
+        23.273,
+        23.323,
+        23.345,
+        23.377,
+        23.445,
+        23.466,
+        23.549,
+        23.611,
+        23.613,
+        23.861,
+        23.868,
+        23.885,
+        23.885,
+        23.906,
+        23.918,
+        23.919,
+        23.933,
+        23.947,
+        24.052,
+        24.064,
+        24.067,
+        24.08,
+        24.103,
+        24.13,
+        24.164,
+        24.249,
+        24.301,
+        24.429,
+        24.515,
+        24.603,
+        24.609,
+        24.7,
+        24.748,
+        24.812,
+        24.826,
+        24.855,
+        24.858,
+        24.954,
+        24.99,
+        25.052,
+        25.097,
+        25.195,
+        25.209,
+        25.241,
+        25.252,
+        25.264,
+        25.298,
+        25.63,
+        25.659,
+        25.767,
+        25.804,
+        25.928,
+        25.976,
+        26.494,
+        26.871,
+        26.886,
+        26.988,
+        26.995,
+        27.004,
+        27.307,
+        27.322,
+        27.352,
+        27.937,
+        28.144,
+        28.165,
+        28.216,
+        28.288,
+        28.331,
+        29.37,
+        29.611,
+        30.6,
+        30.704,
+        31.026,
+        31.552,
+        31.721,
+        32.875,
+        34.001,
+        40.223,
+        49.398
+      ]
+    },
+    "shared_xf_cert_n8": {
+      "p50": 36.185,
+      "p95": 41.136,
+      "p99": 64.287,
+      "max": 64.287,
+      "hz_p50": 27.6,
+      "raw_ms": [
+        33.051,
+        33.153,
+        33.201,
+        33.406,
+        33.819,
+        33.845,
+        33.94,
+        34.083,
+        34.145,
+        34.235,
+        34.3,
+        34.372,
+        34.605,
+        34.697,
+        34.714,
+        34.755,
+        34.817,
+        34.855,
+        34.925,
+        34.957,
+        34.963,
+        34.967,
+        34.986,
+        35.167,
+        35.239,
+        35.314,
+        35.325,
+        35.402,
+        35.444,
+        35.518,
+        35.544,
+        35.582,
+        35.594,
+        35.662,
+        35.675,
+        35.724,
+        35.775,
+        35.874,
+        35.908,
+        35.929,
+        35.965,
+        36.019,
+        36.04,
+        36.126,
+        36.164,
+        36.185,
+        36.2,
+        36.259,
+        36.263,
+        36.266,
+        36.301,
+        36.324,
+        36.383,
+        36.393,
+        36.609,
+        36.667,
+        36.696,
+        36.779,
+        36.799,
+        36.836,
+        36.881,
+        36.978,
+        37.001,
+        37.033,
+        37.043,
+        37.078,
+        37.183,
+        37.206,
+        37.278,
+        37.3,
+        37.411,
+        37.485,
+        37.722,
+        37.923,
+        37.965,
+        38.197,
+        38.588,
+        38.871,
+        39.082,
+        39.099,
+        39.135,
+        39.351,
+        39.505,
+        40.204,
+        40.892,
+        41.136,
+        41.449,
+        42.611,
+        44.687,
+        64.287
+      ]
+    }
+  }
+}

docs/experiment_round34_geom_opt.json

@@ -0,0 +1,630 @@
+{
+  "device": "NVIDIA GeForce RTX 3090",
+  "B": 1,
+  "K": 8,
+  "commit_n": 2,
+  "reps": 80,
+  "design": "5 \u4e2a\u53d8\u4f53\u9010\u8f6e\u4ea4\u66ff\uff08\u5947\u6570\u8f6e\u53cd\u5e8f\uff09\uff1b\u53ea\u6d4b\u53ef\u5b9e\u73b0\u7684\u53e3\u5f84\uff1a\u7279\u5f81/\u63a8\u7406/\u8ba4\u8bc1\u6309\u771f\u5b9e\u987a\u5e8f",
+  "equivalence": {
+    "batch_cert_matrix_equal_to_loop": true,
+    "features_xf_reuse_bitwise_equal": true
+  },
+  "results": {
+    "A_today": {
+      "p50": 33.494,
+      "p95": 46.143,
+      "p99": 64.693,
+      "max": 64.693,
+      "raw_ms": [
+        28.984,
+        29.251,
+        29.361,
+        29.549,
+        29.899,
+        29.944,
+        30.181,
+        30.605,
+        30.624,
+        30.744,
+        30.815,
+        30.838,
+        30.862,
+        30.9,
+        31.024,
+        31.092,
+        31.115,
+        31.262,
+        31.433,
+        31.463,
+        31.52,
+        31.572,
+        31.642,
+        31.741,
+        31.871,
+        31.902,
+        31.924,
+        31.933,
+        31.976,
+        32.001,
+        32.03,
+        32.06,
+        32.285,
+        32.384,
+        32.624,
+        32.825,
+        32.919,
+        33.063,
+        33.168,
+        33.197,
+        33.494,
+        33.589,
+        33.624,
+        33.632,
+        33.666,
+        33.708,
+        33.727,
+        33.735,
+        33.754,
+        33.773,
+        33.869,
+        33.907,
+        34.116,
+        34.219,
+        34.314,
+        34.337,
+        34.897,
+        34.957,
+        35.436,
+        35.469,
+        35.999,
+        36.106,
+        36.523,
+        36.559,
+        36.846,
+        36.866,
+        36.922,
+        37.354,
+        37.375,
+        37.581,
+        37.933,
+        38.026,
+        40.541,
+        40.882,
+        44.161,
+        44.233,
+        46.143,
+        54.047,
+        58.36,
+        64.693
+      ]
+    },
+    "B_xf_reuse": {
+      "p50": 30.122,
+      "p95": 42.349,
+      "p99": 66.368,
+      "max": 66.368,
+      "raw_ms": [
+        26.698,
+        26.958,
+        26.984,
+        27.007,
+        27.018,
+        27.158,
+        27.308,
+        27.369,
+        27.987,
+        28.066,
+        28.219,
+        28.288,
+        28.362,
+        28.517,
+        28.537,
+        28.539,
+        28.593,
+        28.611,
+        28.638,
+        28.735,
+        28.828,
+        28.871,
+        28.922,
+        28.945,
+        28.978,
+        29.088,
+        29.202,
+        29.226,
+        29.226,
+        29.31,
+        29.497,
+        29.506,
+        29.531,
+        29.549,
+        29.649,
+        29.875,
+        29.978,
+        29.992,
+        30.025,
+        30.102,
+        30.122,
+        30.243,
+        30.245,
+        30.263,
+        30.407,
+        30.578,
+        30.665,
+        30.717,
+        30.863,
+        30.884,
+        30.889,
+        30.95,
+        31.023,
+        31.314,
+        31.525,
+        31.537,
+        31.543,
+        31.549,
+        31.62,
+        32.011,
+        32.098,
+        32.168,
+        32.188,
+        32.22,
+        32.578,
+        32.818,
+        32.851,
+        32.929,
+        32.991,
+        33.76,
+        33.805,
+        33.831,
+        33.94,
+        34.124,
+        34.891,
+        39.001,
+        42.349,
+        43.636,
+        62.045,
+        66.368
+      ]
+    },
+    "C_batch_cert": {
+      "p50": 38.048,
+      "p95": 50.171,
+      "p99": 59.86,
+      "max": 59.86,
+      "raw_ms": [
+        33.625,
+        33.736,
+        33.848,
+        34.235,
+        34.49,
+        34.917,
+        34.918,
+        35.108,
+        35.407,
+        35.919,
+        36.023,
+        36.035,
+        36.087,
+        36.263,
+        36.318,
+        36.531,
+        36.627,
+        36.638,
+        36.722,
+        36.778,
+        36.791,
+        36.801,
+        36.816,
+        36.859,
+        36.97,
+        36.998,
+        37.029,
+        37.055,
+        37.057,
+        37.093,
+        37.208,
+        37.318,
+        37.408,
+        37.438,
+        37.463,
+        37.501,
+        37.503,
+        37.605,
+        37.63,
+        37.984,
+        38.048,
+        38.283,
+        38.347,
+        38.524,
+        38.547,
+        38.667,
+        38.816,
+        38.995,
+        39.028,
+        39.34,
+        39.617,
+        39.679,
+        39.845,
+        39.852,
+        39.871,
+        39.894,
+        39.923,
+        40.195,
+        40.514,
+        41.421,
+        41.684,
+        41.804,
+        41.879,
+        41.933,
+        42.011,
+        42.127,
+        42.483,
+        42.551,
+        42.656,
+        43.038,
+        43.248,
+        43.269,
+        45.68,
+        46.514,
+        46.755,
+        48.909,
+        50.171,
+        50.906,
+        55.33,
+        59.86
+      ]
+    },
+    "D_both": {
+      "p50": 34.87,
+      "p95": 39.811,
+      "p99": 48.258,
+      "max": 48.258,
+      "raw_ms": [
+        30.044,
+        30.934,
+        31.005,
+        31.007,
+        31.064,
+        31.492,
+        31.675,
+        31.758,
+        32.114,
+        32.153,
+        32.338,
+        32.374,
+        32.442,
+        32.494,
+        32.681,
+        32.882,
+        32.942,
+        32.991,
+        33.167,
+        33.18,
+        33.302,
+        33.305,
+        33.342,
+        33.449,
+        33.58,
+        33.694,
+        33.739,
+        33.784,
+        33.832,
+        33.835,
+        33.866,
+        33.961,
+        34.037,
+        34.135,
+        34.315,
+        34.395,
+        34.465,
+        34.556,
+        34.656,
+        34.819,
+        34.87,
+        34.98,
+        35.047,
+        35.176,
+        35.19,
+        35.37,
+        35.461,
+        35.528,
+        35.557,
+        35.59,
+        35.643,
+        35.716,
+        35.742,
+        35.804,
+        35.991,
+        36.018,
+        36.063,
+        36.105,
+        36.293,
+        36.312,
+        36.326,
+        36.537,
+        36.541,
+        36.602,
+        36.717,
+        36.934,
+        36.945,
+        37.088,
+        37.141,
+        37.46,
+        37.561,
+        37.608,
+        37.765,
+        37.904,
+        37.974,
+        39.253,
+        39.811,
+        40.076,
+        46.504,
+        48.258
+      ]
+    },
+    "E_both_cont_n8": {
+      "p50": 43.725,
+      "p95": 51.084,
+      "p99": 65.681,
+      "max": 65.681,
+      "raw_ms": [
+        37.733,
+        39.087,
+        39.154,
+        39.238,
+        39.36,
+        39.454,
+        39.514,
+        40.167,
+        40.44,
+        40.837,
+        40.86,
+        40.868,
+        41.237,
+        41.276,
+        41.346,
+        41.408,
+        41.551,
+        41.815,
+        41.824,
+        41.912,
+        41.932,
+        41.973,
+        42.074,
+        42.296,
+        42.341,
+        42.365,
+        42.435,
+        42.505,
+        42.592,
+        42.598,
+        42.637,
+        42.85,
+        43.045,
+        43.048,
+        43.18,
+        43.209,
+        43.359,
+        43.361,
+        43.373,
+        43.411,
+        43.725,
+        43.728,
+        43.787,
+        43.833,
+        43.985,
+        44.133,
+        44.185,
+        44.401,
+        44.458,
+        44.585,
+        44.689,
+        44.819,
+        44.841,
+        44.86,
+        45.081,
+        45.087,
+        45.55,
+        45.701,
+        45.939,
+        46.06,
+        46.128,
+        46.264,
+        46.368,
+        46.43,
+        46.957,
+        47.145,
+        47.231,
+        47.256,
+        47.463,
+        47.637,
+        47.819,
+        47.863,
+        48.382,
+        48.87,
+        49.676,
+        50.884,
+        51.084,
+        51.17,
+        53.859,
+        65.681
+      ]
+    },
+    "F_worst_loop8": {
+      "p50": 80.224,
+      "p95": 88.436,
+      "p99": 89.553,
+      "max": 89.553,
+      "raw_ms": [
+        67.204,
+        70.008,
+        70.436,
+        71.432,
+        72.454,
+        72.503,
+        73.006,
+        73.12,
+        73.492,
+        73.971,
+        74.203,
+        74.358,
+        74.756,
+        74.816,
+        74.893,
+        75.536,
+        75.98,
+        76.049,
+        76.164,
+        76.25,
+        76.547,
+        76.669,
+        76.843,
+        77.268,
+        77.534,
+        77.658,
+        77.872,
+        78.149,
+        78.343,
+        78.377,
+        78.445,
+        78.483,
+        78.86,
+        78.863,
+        78.905,
+        79.048,
+        79.315,
+        79.857,
+        80.035,
+        80.041,
+        80.224,
+        80.602,
+        80.644,
+        80.677,
+        80.848,
+        81.061,
+        81.094,
+        81.155,
+        81.243,
+        81.384,
+        81.628,
+        81.679,
+        81.75,
+        82.117,
+        82.148,
+        82.363,
+        82.444,
+        82.511,
+        82.651,
+        82.828,
+        82.921,
+        83.371,
+        83.592,
+        83.894,
+        83.926,
+        84.386,
+        84.519,
+        84.608,
+        85.06,
+        85.078,
+        85.094,
+        85.26,
+        87.212,
+        87.744,
+        87.847,
+        88.098,
+        88.436,
+        88.461,
+        88.919,
+        89.553
+      ]
+    },
+    "G_worst_batch": {
+      "p50": 37.139,
+      "p95": 46.06,
+      "p99": 52.215,
+      "max": 52.215,
+      "raw_ms": [
+        32.718,
+        33.447,
+        33.569,
+        33.692,
+        33.81,
+        34.236,
+        34.28,
+        34.475,
+        34.496,
+        34.746,
+        35.19,
+        35.392,
+        35.435,
+        35.564,
+        35.569,
+        35.669,
+        35.693,
+        35.734,
+        35.741,
+        35.812,
+        35.828,
+        35.934,
+        36.015,
+        36.036,
+        36.097,
+        36.144,
+        36.197,
+        36.205,
+        36.487,
+        36.637,
+        36.71,
+        36.775,
+        36.79,
+        36.83,
+        36.831,
+        36.885,
+        36.894,
+        36.903,
+        36.947,
+        37.122,
+        37.139,
+        37.154,
+        37.198,
+        37.225,
+        37.249,
+        37.478,
+        37.615,
+        37.617,
+        38.114,
+        38.164,
+        38.288,
+        38.637,
+        38.651,
+        38.724,
+        38.813,
+        39.108,
+        39.15,
+        39.848,
+        39.866,
+        39.946,
+        39.969,
+        39.992,
+        40.04,
+        40.28,
+        40.35,
+        40.697,
+        40.727,
+        40.893,
+        40.972,
+        41.029,
+        41.138,
+        41.479,
+        41.51,
+        41.711,
+        45.036,
+        45.784,
+        46.06,
+        48.597,
+        50.186,
+        52.215
+      ]
+    }
+  }
+}

docs/experiment_round34_precision_ab.json

@@ -0,0 +1,58 @@
+{
+  "device": "NVIDIA GeForce RTX 3090",
+  "torch": "2.10.0+cu130",
+  "K": 8,
+  "steps": 1,
+  "B": 1,
+  "act_scale": 1.0,
+  "inference": {
+    "fp32": {
+      "p50": 7.681,
+      "p95": 9.426,
+      "p99": 10.631,
+      "max": 15.837
+    },
+    "fp16": {
+      "p50": 8.309,
+      "p95": 10.791,
+      "p99": 12.031,
+      "max": 12.322
+    },
+    "bf16": {
+      "p50": 8.577,
+      "p95": 11.057,
+      "p99": 14.174,
+      "max": 17.1
+    }
+  },
+  "cycle": {
+    "fp32": {
+      "p50": 36.507,
+      "p95": 43.688,
+      "p99": 46.568,
+      "max": 50.598
+    },
+    "fp16": {
+      "p50": 36.162,
+      "p95": 50.43,
+      "p99": 55.598,
+      "max": 57.043
+    }
+  },
+  "output_deviation": {
+    "fp16": {
+      "act_max_abs": 0.002096,
+      "act_p99_abs": 0.001059,
+      "deg_max_abs": 0.005,
+      "deg_p99_abs": 0.00219,
+      "would_change_execution_order": false
+    },
+    "bf16": {
+      "act_max_abs": 0.017171,
+      "act_p99_abs": 0.012661,
+      "deg_max_abs": 0.03766,
+      "deg_p99_abs": 0.03008,
+      "would_change_execution_order": true
+    }
+  }
+}

docs/experiment_round34_xf_reuse_check.json

@@ -0,0 +1,70 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round34_xf_reuse_check.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 0,
+    "gap_margin": 0.0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.57421875,
+    "collision_free_success": 0.57421875,
+    "time_ok_1p3": 0.390625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.4449527263641357,
+    "time_mean_s": 2.3076188564300537,
+    "brakes_mean": 19.25,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.20997506380081177,
+    "err_fail_med": 229.68746948242188,
+    "err_success_med": 0.18899640440940857,
+    "dls_share": 0.625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.6600004434585571,
+    "speed_ratio_p95": 1.6600035429000854,
+    "speed_ratio_max": 1.660003662109375,
+    "over_vmax_rate": 1.0,
+    "accel_ratio_p95": 1.5835107564926147,
+    "accel_ratio_max": 1.6600004434585571
+  },
+  "simulation_wall_s": 317.96790459999465
+}

docs/experiment_round35_pipeline_gaps.json

@@ -0,0 +1,23 @@
+{
+  "device": "NVIDIA GeForce RTX 3090",
+  "K": 8,
+  "commit_n": 2,
+  "design": "\u4e24\u6b21\u89c4\u5212\u4e3a\u4e00\u4e2a\u5355\u5143\uff0c\u5947\u6570\u8f6e\u53cd\u5e8f\uff1b\u6d41\u6c34\u7ebf\u90a3\u6b21\u7684\u8ba4\u8bc1\u8c03\u7528\u88ab\u8feb\u7528 cap=GAP_CAP(\u7279\u5f81\u9700\u8981)\uff0c\u8fd9\u662f\u590d\u7528\u7684\u771f\u5b9e\u4ee3\u4ef7\uff0c\u4e0d\u662f\u514d\u8d39",
+  "equivalence": {
+    "cached_gap_row_bitwise_equal": false,
+    "features_bitwise_equal": false,
+    "max_row_diff_mm": 6.29425048828125e-05
+  },
+  "fresh": {
+    "p50_per_plan": 34.779,
+    "p95": 39.271,
+    "p99": 40.98,
+    "reps": 60
+  },
+  "pipelined": {
+    "p50_per_plan": 31.655,
+    "p95": 34.483,
+    "p99": 38.208,
+    "reps": 60
+  }
+}

docs/experiment_round36_kmf_e1.json

@@ -0,0 +1,337 @@
+{
+ "checkpoint": "ml/runs/base_dag2_dag1/best.pt",
+ "n": 256,
+ "nref_steps": 128,
+ "reference": "瞬时场 r=t 的流图（与损失无关）",
+ "per_rt": [
+  {
+   "r": 0.0,
+   "t": 0.05,
+   "span": 0.05,
+   "norm_u_true": 4.152298927307129,
+   "norm_dudt": 156.59774780273438,
+   "mf_abs": 2.629317283630371,
+   "mf_rel": 0.6185885667800903,
+   "mf_cos_true": -0.2586098909378052,
+   "mf_cos_eps": -0.08821636438369751,
+   "kmf_sq_abs": 3.808063507080078,
+   "kmf_lin_abs": 2.105757713317871,
+   "kmf_lin_rel": 0.4867732524871826,
+   "kmf_lin_cos_eps": -0.19339147210121155,
+   "mid_only_abs": 2.467759370803833,
+   "midstate_abs": 2.475457191467285,
+   "midstate_rel": 0.5789506435394287,
+   "heun_abs": 2.465583324432373,
+   "heun_rel": 0.5782649517059326,
+   "d1_norm": 157.7960968017578,
+   "d2_norm": 150.50238037109375,
+   "d3_norm": 153.35733032226562,
+   "net_abs": 2.7900571823120117,
+   "net_rel": 0.686574399471283
+  },
+  {
+   "r": 0.0,
+   "t": 0.1,
+   "span": 0.1,
+   "norm_u_true": 6.895779609680176,
+   "norm_dudt": 25.08617401123047,
+   "mf_abs": 2.9073116779327393,
+   "mf_rel": 0.4195072650909424,
+   "mf_cos_true": 0.7109781503677368,
+   "mf_cos_eps": 0.6336730718612671,
+   "kmf_sq_abs": 1.5212182998657227,
+   "kmf_lin_abs": 5.014378070831299,
+   "kmf_lin_rel": 0.7285130023956299,
+   "kmf_lin_cos_eps": -0.707574188709259,
+   "mid_only_abs": 11.992480278015137,
+   "midstate_abs": 2.770589828491211,
+   "midstate_rel": 0.39942666888237,
+   "heun_abs": 1.8834517002105713,
+   "heun_rel": 0.2704611122608185,
+   "d1_norm": 24.996402740478516,
+   "d2_norm": 156.59774780273438,
+   "d3_norm": 59.145477294921875,
+   "net_abs": 3.2193822860717773,
+   "net_rel": 0.4677886366844177
+  },
+  {
+   "r": 0.0,
+   "t": 0.2,
+   "span": 0.2,
+   "norm_u_true": 8.788705825805664,
+   "norm_dudt": 6.331303596496582,
+   "mf_abs": 2.572011947631836,
+   "mf_rel": 0.29417163133621216,
+   "mf_cos_true": 0.8289472460746765,
+   "mf_cos_eps": 0.795153021812439,
+   "kmf_sq_abs": 1.7290871143341064,
+   "kmf_lin_abs": 1.926774501800537,
+   "kmf_lin_rel": 0.2203272432088852,
+   "kmf_lin_cos_eps": 0.08155926316976547,
+   "mid_only_abs": 3.628005266189575,
+   "midstate_abs": 3.9586117267608643,
+   "midstate_rel": 0.4504213333129883,
+   "heun_abs": 1.7300267219543457,
+   "heun_rel": 0.19759969413280487,
+   "d1_norm": 4.81895637512207,
+   "d2_norm": 25.08617401123047,
+   "d3_norm": 28.122135162353516,
+   "net_abs": 1.9169456958770752,
+   "net_rel": 0.2185157686471939
+  },
+  {
+   "r": 0.0,
+   "t": 0.3,
+   "span": 0.3,
+   "norm_u_true": 9.446111679077148,
+   "norm_dudt": 4.131267547607422,
+   "mf_abs": 2.354419231414795,
+   "mf_rel": 0.2489214837551117,
+   "mf_cos_true": 0.8489015102386475,
+   "mf_cos_eps": 0.8029109239578247,
+   "kmf_sq_abs": 1.4227628707885742,
+   "kmf_lin_abs": 1.530599594116211,
+   "kmf_lin_rel": 0.16132134199142456,
+   "kmf_lin_cos_eps": 0.09631951153278351,
+   "mid_only_abs": 2.57088565826416,
+   "midstate_abs": 4.304656982421875,
+   "midstate_rel": 0.4555080831050873,
+   "heun_abs": 1.8236483335494995,
+   "heun_rel": 0.1927146017551422,
+   "d1_norm": 3.2171449661254883,
+   "d2_norm": 11.126049041748047,
+   "d3_norm": 18.29114532470703,
+   "net_abs": 1.3471484184265137,
+   "net_rel": 0.14259807765483856
+  },
+  {
+   "r": 0.0,
+   "t": 0.4,
+   "span": 0.4,
+   "norm_u_true": 9.78399658203125,
+   "norm_dudt": 2.373415946960449,
+   "mf_abs": 1.7019946575164795,
+   "mf_rel": 0.17237165570259094,
+   "mf_cos_true": 0.8364542722702026,
+   "mf_cos_eps": 0.80040043592453,
+   "kmf_sq_abs": 1.4873857498168945,
+   "kmf_lin_abs": 1.8382024765014648,
+   "kmf_lin_rel": 0.18990299105644226,
+   "kmf_lin_cos_eps": 0.6340612173080444,
+   "mid_only_abs": 2.7560770511627197,
+   "midstate_abs": 4.497618675231934,
+   "midstate_rel": 0.4598080515861511,
+   "heun_abs": 2.0347068309783936,
+   "heun_rel": 0.20750463008880615,
+   "d1_norm": 2.2074503898620605,
+   "d2_norm": 6.331303596496582,
+   "d3_norm": 13.588470458984375,
+   "net_abs": 0.9387868642807007,
+   "net_rel": 0.09583067893981934
+  },
+  {
+   "r": 0.0,
+   "t": 0.5,
+   "span": 0.5,
+   "norm_u_true": 9.98532772064209,
+   "norm_dudt": 2.658200263977051,
+   "mf_abs": 0.8106092214584351,
+   "mf_rel": 0.0805797278881073,
+   "mf_cos_true": -0.3406544327735901,
+   "mf_cos_eps": -0.19956746697425842,
+   "kmf_sq_abs": 1.2309105396270752,
+   "kmf_lin_abs": 1.5382835865020752,
+   "kmf_lin_rel": 0.15414834022521973,
+   "kmf_lin_cos_eps": 0.5139043927192688,
+   "mid_only_abs": 2.382789134979248,
+   "midstate_abs": 5.330208778381348,
+   "midstate_rel": 0.5345790982246399,
+   "heun_abs": 2.8818538188934326,
+   "heun_rel": 0.28885430097579956,
+   "d1_norm": 1.423255443572998,
+   "d2_norm": 4.9012861251831055,
+   "d3_norm": 12.108942031860352,
+   "net_abs": 0.8207994103431702,
+   "net_rel": 0.08221843093633652
+  },
+  {
+   "r": 0.0,
+   "t": 0.6,
+   "span": 0.6,
+   "norm_u_true": 10.118559837341309,
+   "norm_dudt": 4.046132564544678,
+   "mf_abs": 3.044166088104248,
+   "mf_rel": 0.3003576099872589,
+   "mf_cos_true": 0.9172172546386719,
+   "mf_cos_eps": 0.841435432434082,
+   "kmf_sq_abs": 1.3112163543701172,
+   "kmf_lin_abs": 1.7851487398147583,
+   "kmf_lin_rel": 0.17629936337471008,
+   "kmf_lin_cos_eps": 0.7054632306098938,
+   "mid_only_abs": 2.974583625793457,
+   "midstate_abs": 3.8119943141937256,
+   "midstate_rel": 0.37495750188827515,
+   "heun_abs": 0.9826163649559021,
+   "heun_rel": 0.09552386403083801,
+   "d1_norm": 1.021875023841858,
+   "d2_norm": 4.131267547607422,
+   "d3_norm": 7.380629539489746,
+   "net_abs": 0.5622822046279907,
+   "net_rel": 0.0555611215531826
+  },
+  {
+   "r": 0.0,
+   "t": 0.7,
+   "span": 0.7,
+   "norm_u_true": 10.210578918457031,
+   "norm_dudt": 1.5618565082550049,
+   "mf_abs": 0.7238559722900391,
+   "mf_rel": 0.0707138329744339,
+   "mf_cos_true": 0.01889939233660698,
+   "mf_cos_eps": 0.0709509551525116,
+   "kmf_sq_abs": 1.1261842250823975,
+   "kmf_lin_abs": 1.562591552734375,
+   "kmf_lin_rel": 0.15248213708400726,
+   "kmf_lin_cos_eps": 0.6589925289154053,
+   "mid_only_abs": 2.7245984077453613,
+   "midstate_abs": 5.200009346008301,
+   "midstate_rel": 0.509149432182312,
+   "heun_abs": 2.7610363960266113,
+   "heun_rel": 0.27028360962867737,
+   "d1_norm": 0.9923672080039978,
+   "d2_norm": 3.2669692039489746,
+   "d3_norm": 8.216880798339844,
+   "net_abs": 0.6127126216888428,
+   "net_rel": 0.06004098802804947
+  },
+  {
+   "r": 0.0,
+   "t": 0.8,
+   "span": 0.8,
+   "norm_u_true": 10.280738830566406,
+   "norm_dudt": 3.559027671813965,
+   "mf_abs": 3.139580726623535,
+   "mf_rel": 0.3054307699203491,
+   "mf_cos_true": 0.841235876083374,
+   "mf_cos_eps": 0.7405093908309937,
+   "kmf_sq_abs": 0.9379472732543945,
+   "kmf_lin_abs": 1.278116226196289,
+   "kmf_lin_rel": 0.12322138249874115,
+   "kmf_lin_cos_eps": 0.6785725951194763,
+   "mid_only_abs": 2.1469192504882812,
+   "midstate_abs": 3.9412341117858887,
+   "midstate_rel": 0.3837699890136719,
+   "heun_abs": 1.2848330736160278,
+   "heun_rel": 0.12474364042282104,
+   "d1_norm": 0.6407566070556641,
+   "d2_norm": 2.373415946960449,
+   "d3_norm": 5.565019130706787,
+   "net_abs": 0.544069766998291,
+   "net_rel": 0.05294711887836456
+  },
+  {
+   "r": 0.0,
+   "t": 0.9,
+   "span": 0.9,
+   "norm_u_true": 10.333314895629883,
+   "norm_dudt": 17.87816619873047,
+   "mf_abs": 16.365367889404297,
+   "mf_rel": 1.5846014022827148,
+   "mf_cos_true": 0.949444055557251,
+   "mf_cos_eps": 0.8185906410217285,
+   "kmf_sq_abs": 0.7964116334915161,
+   "kmf_lin_abs": 1.5619468688964844,
+   "kmf_lin_rel": 0.14914938807487488,
+   "kmf_lin_cos_eps": -0.682485818862915,
+   "mid_only_abs": 2.9376566410064697,
+   "midstate_abs": 4.5689287185668945,
+   "midstate_rel": 0.44398587942123413,
+   "heun_abs": 10.097332000732422,
+   "heun_rel": 0.9785149693489075,
+   "d1_norm": 0.9012945890426636,
+   "d2_norm": 3.5273427963256836,
+   "d3_norm": 4.809117794036865,
+   "net_abs": 0.19002258777618408,
+   "net_rel": 0.018513960763812065
+  },
+  {
+   "r": 0.0,
+   "t": 0.95,
+   "span": 0.95,
+   "norm_u_true": 10.35672664642334,
+   "norm_dudt": 64.29971313476562,
+   "mf_abs": 61.504981994628906,
+   "mf_rel": 5.942716598510742,
+   "mf_cos_true": 0.9920964241027832,
+   "mf_cos_eps": 0.938286304473877,
+   "kmf_sq_abs": 0.6677861213684082,
+   "kmf_lin_abs": 1.1961216926574707,
+   "kmf_lin_rel": 0.1146148070693016,
+   "kmf_lin_cos_eps": -0.46628355979919434,
+   "mid_only_abs": 2.5505270957946777,
+   "midstate_abs": 26.227705001831055,
+   "midstate_rel": 2.537191390991211,
+   "heun_abs": 43.71785354614258,
+   "heun_rel": 4.225504398345947,
+   "d1_norm": 0.9435088634490967,
+   "d2_norm": 2.9939303398132324,
+   "d3_norm": 27.214248657226562,
+   "net_abs": 2.2616631984710693,
+   "net_rel": 0.2185918092727661
+  },
+  {
+   "r": 0.0,
+   "t": 1.0,
+   "span": 1.0,
+   "norm_u_true": 10.378573417663574,
+   "norm_dudt": 4.240537643432617,
+   "mf_abs": 3.8513689041137695,
+   "mf_rel": 0.3713422417640686,
+   "mf_cos_true": -0.9887999892234802,
+   "mf_cos_eps": -0.916346549987793,
+   "kmf_sq_abs": 7.18745231628418,
+   "kmf_lin_abs": 14.660856246948242,
+   "kmf_lin_rel": 1.412999153137207,
+   "kmf_lin_cos_eps": -0.8949825167655945,
+   "mid_only_abs": 2.364427328109741,
+   "midstate_abs": 4.872905731201172,
+   "midstate_rel": 0.4697429835796356,
+   "heun_abs": 4.360708236694336,
+   "heun_rel": 0.4204045832157135,
+   "d1_norm": 28.04054832458496,
+   "d2_norm": 2.658200263977051,
+   "d3_norm": 5.2787017822265625,
+   "net_abs": 3.9062576293945312,
+   "net_rel": 0.3766472339630127
+  }
+ ],
+ "ref_discretization": {
+  "norm_128": 10.378573417663574,
+  "norm_512": 10.360570907592773,
+  "diff_mean": 0.018837258219718933,
+  "diff_rel": 0.0018179214093834162
+ },
+ "gate_corner": {
+  "mf": 3.8513689041137695,
+  "kmf_sq": 7.18745231628418,
+  "kmf_lin": 14.660856246948242,
+  "mid_only": 2.364427328109741,
+  "midstate": 4.872905731201172,
+  "heun": 4.360708236694336,
+  "net_1step": 3.9062576293945312,
+  "ratio_lin_over_mf": 3.8066611150358813,
+  "ratio_mid_over_mf": 0.6139186837137884,
+  "ratio_midstate_over_mf": 1.265239932221571,
+  "ratio_heun_over_mf": 1.132248908183406
+ },
+ "gate_worst_accel": {
+  "t": 0.05,
+  "dudt": 156.59774780273438,
+  "mf": 2.629317283630371,
+  "kmf_sq": 3.808063507080078,
+  "kmf_lin": 2.105757713317871,
+  "mid_only": 2.467759370803833,
+  "ratio_lin_over_mf": 0.8008762298973645
+ },
+ "wall_s": 13.092682200003765
+}

docs/experiment_round36_kmf_e1b.json

@@ -0,0 +1,97 @@
+{
+ "n": 512,
+ "steps": {
+  "2": {
+   "vs_teacher_l2": 0.4342995882034302,
+   "vs_32step_l2": 0.30502474308013916,
+   "vs_1step_l2": 4.2227020263671875
+  },
+  "3": {
+   "vs_teacher_l2": 0.43872418999671936,
+   "vs_32step_l2": 0.29536527395248413,
+   "vs_1step_l2": 4.216776371002197
+  },
+  "4": {
+   "vs_teacher_l2": 0.42935511469841003,
+   "vs_32step_l2": 0.29940515756607056,
+   "vs_1step_l2": 4.232732772827148
+  },
+  "8": {
+   "vs_teacher_l2": 0.4444255232810974,
+   "vs_32step_l2": 0.2452448457479477,
+   "vs_1step_l2": 4.195967197418213
+  },
+  "16": {
+   "vs_teacher_l2": 0.5166380405426025,
+   "vs_32step_l2": 0.12673738598823547,
+   "vs_1step_l2": 4.085569381713867
+  },
+  "32": {
+   "vs_teacher_l2": 0.6008902192115784,
+   "vs_32step_l2": 0.0,
+   "vs_1step_l2": 3.970090866088867
+  }
+ },
+ "vs_teacher": {
+  "a1_1step": {
+   "l2_mean": 4.325636386871338,
+   "l2_p95": 4.9564337730407715,
+   "per_axis_norm": [
+    0.3975203335285187,
+    0.36649173498153687,
+    0.3425973951816559,
+    0.3447648882865906,
+    0.3368167281150818,
+    0.32962310314178467
+   ]
+  },
+  "aU_32step_rneq_t": {
+   "l2_mean": 0.6008902192115784,
+   "l2_p95": 1.9733507633209229,
+   "per_axis_norm": [
+    0.05318881571292877,
+    0.04853545129299164,
+    0.04775547981262207,
+    0.04531646892428398,
+    0.0381026417016983,
+    0.034751877188682556
+   ]
+  },
+  "aF_flow_map_r_eq_t": {
+   "l2_mean": 0.6651337146759033,
+   "l2_p95": 1.9794291257858276,
+   "per_axis_norm": [
+    0.06039007008075714,
+    0.05440700799226761,
+    0.05297258496284485,
+    0.05065540224313736,
+    0.04315900802612305,
+    0.03988185524940491
+   ]
+  }
+ },
+ "mutual": [
+  {
+   "l2_mean": 3.970090866088867,
+   "pair": [
+    "a1",
+    "aU_32"
+   ]
+  },
+  {
+   "l2_mean": 3.891127109527588,
+   "pair": [
+    "a1",
+    "aF_flow"
+   ]
+  },
+  {
+   "l2_mean": 0.08254596590995789,
+   "pair": [
+    "aU_32",
+    "aF_flow"
+   ]
+  }
+ ],
+ "note": "动作块是 ±1 归一化的 16×6；l2_mean 无单位权重，越小越贴老师"
+}

docs/experiment_round36_plan_r_hist.json

@@ -0,0 +1,50 @@
+{
+ "empty/commit2/inc_cap0": {
+  "plans": 175,
+  "r_hist": {
+   "0": 175
+  },
+  "frac_r0": 1.0,
+  "save_ms_per_plan_avg": 3.631,
+  "success": 0.46875,
+  "brakes_mean": 24.0625,
+  "wall_s": 54.74565179999627
+ },
+ "r3/commit3/inc_cap0": {
+  "plans": 118,
+  "r_hist": {
+   "0": 1,
+   "2": 1,
+   "3": 116
+  },
+  "frac_r0": 0.00847457627118644,
+  "save_ms_per_plan_avg": 0.030771186440677965,
+  "success": 0.4375,
+  "brakes_mean": 18.03125,
+  "wall_s": 49.8360657000012
+ },
+ "r3/commit3/inc_cap1": {
+  "plans": 118,
+  "r_hist": {
+   "0": 1,
+   "2": 1,
+   "3": 116
+  },
+  "frac_r0": 0.00847457627118644,
+  "save_ms_per_plan_avg": 0.030771186440677965,
+  "success": 0.40625,
+  "brakes_mean": 16.9375,
+  "wall_s": 49.97621490000165
+ },
+ "empty/commit3/inc_cap1": {
+  "plans": 117,
+  "r_hist": {
+   "0": 117
+  },
+  "frac_r0": 1.0,
+  "save_ms_per_plan_avg": 3.631,
+  "success": 0.4375,
+  "brakes_mean": 15.0625,
+  "wall_s": 43.831084999997984
+ }
+}

docs/experiment_round36_steps2_cap0.json

@@ -0,0 +1,73 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round36_steps2_cap0_rows.json",
+    "out": "docs/experiment_round36_steps2_cap0.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 0,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 2,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.42578125,
+    "collision_free_success": 0.42578125,
+    "time_ok_1p3": 0.35546875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 1.5711719989776611,
+    "time_mean_s": 1.6664220094680786,
+    "brakes_mean": 86.30859375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.29959768056869507,
+    "err_fail_med": 405.61505126953125,
+    "err_success_med": 0.24314510822296143,
+    "dls_share": 0.453125,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.6600004434585571,
+    "speed_ratio_p95": 1.6600035429000854,
+    "speed_ratio_max": 1.660003662109375,
+    "over_vmax_rate": 0.91015625,
+    "accel_ratio_p95": 1.0188239812850952,
+    "accel_ratio_max": 1.4423871040344238
+  },
+  "simulation_wall_s": 304.77464890001283
+}

docs/experiment_round36_steps2_cap1.json

@@ -0,0 +1,73 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round36_steps2_cap1_rows.json",
+    "out": "docs/experiment_round36_steps2_cap1.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 2,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.39453125,
+    "collision_free_success": 0.39453125,
+    "time_ok_1p3": 0.296875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 1.8436782360076904,
+    "time_mean_s": 1.9299010038375854,
+    "brakes_mean": 86.15625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.384182870388031,
+    "err_fail_med": 376.4237976074219,
+    "err_success_med": 0.2933140993118286,
+    "dls_share": 0.4453125,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000038146972656,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 0.8507659435272217,
+    "accel_ratio_max": 0.9923425912857056
+  },
+  "simulation_wall_s": 319.5858334999939
+}

docs/experiment_round36_steps4_cap1.json

@@ -0,0 +1,73 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round36_steps4_cap1_rows.json",
+    "out": "docs/experiment_round36_steps4_cap1.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 4,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.40625,
+    "collision_free_success": 0.40625,
+    "time_ok_1p3": 0.3046875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 1.8817013502120972,
+    "time_mean_s": 1.9080770015716553,
+    "brakes_mean": 79.65234375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.3844277262687683,
+    "err_fail_med": 342.1692810058594,
+    "err_success_med": 0.1346309632062912,
+    "dls_share": 0.4609375,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000053644180298,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 0.914295494556427,
+    "accel_ratio_max": 1.0000005960464478
+  },
+  "simulation_wall_s": 347.79689209999924
+}

docs/experiment_round36_steps8_cap1.json

@@ -0,0 +1,73 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round36_steps8_cap1_rows.json",
+    "out": "docs/experiment_round36_steps8_cap1.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 8,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.41015625,
+    "collision_free_success": 0.41015625,
+    "time_ok_1p3": 0.32421875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 1.8700296878814697,
+    "time_mean_s": 1.9293333292007446,
+    "brakes_mean": 79.69921875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.3575941324234009,
+    "err_fail_med": 330.5984191894531,
+    "err_success_med": 0.2526792883872986,
+    "dls_share": 0.46875,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000053644180298,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 0.9362331628799438,
+    "accel_ratio_max": 1.0000005960464478
+  },
+  "simulation_wall_s": 316.32940030000464
+}

docs/experiment_round36_time_audit.json

@@ -0,0 +1,106 @@
+[
+ {
+  "file": "ml/runs/round33_cap.pt",
+  "n": 256,
+  "all": {
+   "n": 256,
+   "mean_window_s": 3.73765625,
+   "zero_share": 0.20881234062121148,
+   "low_share": 0.0812884076752644,
+   "mid_share": 0.27427783119434807,
+   "hi_share": 0.435621420509176,
+   "zero_s_per_task": 0.78046875,
+   "runs_per_task": 18.5078125,
+   "run_len_mean": 2.1084845926551288,
+   "run_len_max": 41
+  },
+  "arrived_time_ok": {
+   "n": 68,
+   "mean_window_s": 1.9370588235294117,
+   "zero_share": 0.013817187974491345,
+   "low_share": 0.033404190707561496,
+   "mid_share": 0.17901609474643185,
+   "hi_share": 0.7737625265715153,
+   "zero_s_per_task": 0.02676470588235294,
+   "runs_per_task": 0.7647058823529411,
+   "run_len_mean": 1.75,
+   "run_len_max": 4
+  },
+  "arrived_slow": {
+   "n": 66,
+   "mean_window_s": 3.052727272727273,
+   "zero_share": 0.049632717887631524,
+   "low_share": 0.06293428628151677,
+   "mid_share": 0.22553107008139764,
+   "hi_share": 0.6619019257494541,
+   "zero_s_per_task": 0.15151515151515152,
+   "runs_per_task": 4.590909090909091,
+   "run_len_mean": 1.6501650165016502,
+   "run_len_max": 7
+  },
+  "not_arrived": {
+   "n": 122,
+   "mean_window_s": 5.111803278688525,
+   "zero_share": 0.3014238984029248,
+   "low_share": 0.09733179398370856,
+   "mid_share": 0.31014687961003146,
+   "hi_share": 0.29109742800333527,
+   "zero_s_per_task": 1.5408196721311478,
+   "runs_per_task": 35.92622950819672,
+   "run_len_mean": 2.144421629021218,
+   "run_len_max": 41
+  }
+ },
+ {
+  "file": "ml/runs/round33_base.pt",
+  "n": 256,
+  "all": {
+   "n": 256,
+   "mean_window_s": 3.482578125,
+   "zero_share": 0.22394957040626332,
+   "low_share": 0.08179105816901093,
+   "mid_share": 0.2871435942302084,
+   "hi_share": 0.4071157771945174,
+   "zero_s_per_task": 0.779921875,
+   "runs_per_task": 18.7421875,
+   "run_len_mean": 2.0806586077532305,
+   "run_len_max": 21
+  },
+  "arrived_time_ok": {
+   "n": 100,
+   "mean_window_s": 1.8818000000000001,
+   "zero_share": 0.019236900839621637,
+   "low_share": 0.039642895100435754,
+   "mid_share": 0.19353810181740885,
+   "hi_share": 0.7475821022425337,
+   "zero_s_per_task": 0.0362,
+   "runs_per_task": 1.2,
+   "run_len_mean": 1.5083333333333333,
+   "run_len_max": 5
+  },
+  "arrived_slow": {
+   "n": 47,
+   "mean_window_s": 3.2136170212765958,
+   "zero_share": 0.10474046610169492,
+   "low_share": 0.07216631355932204,
+   "mid_share": 0.25529661016949157,
+   "hi_share": 0.5677966101694916,
+   "zero_s_per_task": 0.33659574468085107,
+   "runs_per_task": 10.063829787234043,
+   "run_len_mean": 1.6723044397463003,
+   "run_len_max": 7
+  },
+  "not_arrived": {
+   "n": 109,
+   "mean_window_s": 5.067155963302753,
+   "zero_share": 0.3262963499420625,
+   "low_share": 0.09878331402085747,
+   "mid_share": 0.32774478563151793,
+   "hi_share": 0.247175550405562,
+   "zero_s_per_task": 1.6533944954128441,
+   "runs_per_task": 38.57798165137615,
+   "run_len_mean": 2.142925089179548,
+   "run_len_max": 21
+  }
+ }
+]

docs/experiment_round38_backoff0.json

@@ -0,0 +1,82 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round38_backoff0.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5234375,
+    "collision_free_success": 0.5234375,
+    "time_ok_1p3": 0.265625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.615795850753784,
+    "time_mean_s": 2.486567258834839,
+    "brakes_mean": 19.5546875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.27920001745224,
+    "err_fail_med": 204.71360778808594,
+    "err_success_med": 0.21751396358013153,
+    "dls_share": 0.61328125,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.24741201102733612,
+    "reverse_mean": 0.25991496443748474,
+    "reverse_moving_p50": 0.32581019401550293,
+    "moving_share_p50": 0.9587156176567078,
+    "brake_rate_mean": 0.10028044879436493,
+    "brake_rate_p50": 0.0051282052882015705,
+    "n_plans_mean": 195.0
+  },
+  "simulation_wall_s": 319.19268320000265
+}

docs/experiment_round38_backoff1.json

@@ -0,0 +1,84 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round38_backoff1.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 1,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.52734375,
+    "collision_free_success": 0.52734375,
+    "time_ok_1p3": 0.26953125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.615795850753784,
+    "time_mean_s": 2.479703664779663,
+    "brakes_mean": 18.77734375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.27920001745224,
+    "err_fail_med": 212.4702606201172,
+    "err_success_med": 0.20242473483085632,
+    "dls_share": 0.6171875,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.2482638955116272,
+    "reverse_mean": 0.2603996992111206,
+    "reverse_moving_p50": 0.3252427279949188,
+    "moving_share_p50": 0.9640287756919861,
+    "brake_rate_mean": 0.09629407525062561,
+    "brake_rate_p50": 0.0051282052882015705,
+    "n_plans_mean": 195.0,
+    "backoff_mean": 1.57421875,
+    "reject_all_mean": 20.3515625
+  },
+  "simulation_wall_s": 334.8870786000043
+}

docs/experiment_round38_backoff2.json

@@ -0,0 +1,84 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round38_backoff2.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 2,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5234375,
+    "collision_free_success": 0.5234375,
+    "time_ok_1p3": 0.26171875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.5341429710388184,
+    "time_mean_s": 2.4558210372924805,
+    "brakes_mean": 18.1171875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.274954229593277,
+    "err_fail_med": 207.97683715820312,
+    "err_success_med": 0.19538111984729767,
+    "dls_share": 0.60546875,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.24827586114406586,
+    "reverse_mean": 0.2605503499507904,
+    "reverse_moving_p50": 0.32248520851135254,
+    "moving_share_p50": 0.9672130942344666,
+    "brake_rate_mean": 0.09290865808725357,
+    "brake_rate_p50": 0.0051282052882015705,
+    "n_plans_mean": 195.0,
+    "backoff_mean": 4.99609375,
+    "reject_all_mean": 23.11328125
+  },
+  "simulation_wall_s": 369.2254191000102
+}

docs/experiment_round38_backoff3.json

@@ -0,0 +1,84 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round38_backoff3.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 3,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.515625,
+    "collision_free_success": 0.515625,
+    "time_ok_1p3": 0.2578125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.529658794403076,
+    "time_mean_s": 2.4427273273468018,
+    "brakes_mean": 16.48046875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2868920564651489,
+    "err_fail_med": 208.72935485839844,
+    "err_success_med": 0.20242473483085632,
+    "dls_share": 0.60546875,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.25,
+    "reverse_mean": 0.26120537519454956,
+    "reverse_moving_p50": 0.3236245810985565,
+    "moving_share_p50": 0.9672130942344666,
+    "brake_rate_mean": 0.08451522886753082,
+    "brake_rate_p50": 0.0051282052882015705,
+    "n_plans_mean": 195.0,
+    "backoff_mean": 9.703125,
+    "reject_all_mean": 26.18359375
+  },
+  "simulation_wall_s": 416.57447530000354
+}

docs/experiment_round38_cand_div.json

@@ -0,0 +1,22 @@
+{
+ "K": 8,
+ "n": 256,
+ "steps1": {
+  "cand_spread": 0.8518,
+  "first_step_norm": 1.1565,
+  "clamp_saturated_rate": 0.0516,
+  "pairwise_mean": 5.3647
+ },
+ "steps2": {
+  "cand_spread": 0.0245,
+  "first_step_norm": 1.0956,
+  "clamp_saturated_rate": 0.0385,
+  "pairwise_mean": 0.3024
+ },
+ "steps4": {
+  "cand_spread": 0.0211,
+  "first_step_norm": 1.097,
+  "clamp_saturated_rate": 0.0689,
+  "pairwise_mean": 0.2861
+ }
+}

docs/experiment_round38_fail_taxonomy.json

@@ -0,0 +1,19 @@
+[
+ {
+  "label": "steps=1 cap=0",
+  "n": 256,
+  "arrived": 147,
+  "failed": 109,
+  "buckets": {
+   "姿态没到位(位置<5mm)": 3,
+   "位置 5~20mm": 5,
+   "位置 20~100mm": 15,
+   "位置 >100mm 且刹车>30(被挡)": 63,
+   "位置 >100mm 且刹车<=30(方向/绕圈)": 23,
+   "起步失败": 0
+  },
+  "failed_e_p50": 255.333598802122,
+  "failed_oe_p50": 11.016554443385107,
+  "failed_still_share_p50": 0.3722222149372101
+ }
+]

docs/experiment_round38_steps_cost.json

@@ -0,0 +1,611 @@
+{
+ "device": "NVIDIA GeForce RTX 3090",
+ "B": 1,
+ "K": 8,
+ "rounds": 24,
+ "per_round": {
+  "1": [
+   {
+    "p50": 9.198,
+    "p95": 11.654,
+    "p99": 12.236,
+    "max": 12.236
+   },
+   {
+    "p50": 9.211,
+    "p95": 13.105,
+    "p99": 13.765,
+    "max": 13.765
+   },
+   {
+    "p50": 9.188,
+    "p95": 12.363,
+    "p99": 15.901,
+    "max": 15.901
+   },
+   {
+    "p50": 8.894,
+    "p95": 10.426,
+    "p99": 11.641,
+    "max": 11.641
+   },
+   {
+    "p50": 8.968,
+    "p95": 11.664,
+    "p99": 11.874,
+    "max": 11.874
+   },
+   {
+    "p50": 9.456,
+    "p95": 13.502,
+    "p99": 13.951,
+    "max": 13.951
+   },
+   {
+    "p50": 8.683,
+    "p95": 10.991,
+    "p99": 11.978,
+    "max": 11.978
+   },
+   {
+    "p50": 9.599,
+    "p95": 13.894,
+    "p99": 17.102,
+    "max": 17.102
+   },
+   {
+    "p50": 9.405,
+    "p95": 13.07,
+    "p99": 13.111,
+    "max": 13.111
+   },
+   {
+    "p50": 8.947,
+    "p95": 10.154,
+    "p99": 10.405,
+    "max": 10.405
+   },
+   {
+    "p50": 9.019,
+    "p95": 10.835,
+    "p99": 12.272,
+    "max": 12.272
+   },
+   {
+    "p50": 8.824,
+    "p95": 13.516,
+    "p99": 15.963,
+    "max": 15.963
+   },
+   {
+    "p50": 8.819,
+    "p95": 11.004,
+    "p99": 11.308,
+    "max": 11.308
+   },
+   {
+    "p50": 8.777,
+    "p95": 10.7,
+    "p99": 11.144,
+    "max": 11.144
+   },
+   {
+    "p50": 8.926,
+    "p95": 11.911,
+    "p99": 12.171,
+    "max": 12.171
+   },
+   {
+    "p50": 8.924,
+    "p95": 10.88,
+    "p99": 11.94,
+    "max": 11.94
+   },
+   {
+    "p50": 8.508,
+    "p95": 9.918,
+    "p99": 11.57,
+    "max": 11.57
+   },
+   {
+    "p50": 9.351,
+    "p95": 17.795,
+    "p99": 17.948,
+    "max": 17.948
+   },
+   {
+    "p50": 8.779,
+    "p95": 10.188,
+    "p99": 11.614,
+    "max": 11.614
+   },
+   {
+    "p50": 9.389,
+    "p95": 19.998,
+    "p99": 25.582,
+    "max": 25.582
+   },
+   {
+    "p50": 9.344,
+    "p95": 15.293,
+    "p99": 16.308,
+    "max": 16.308
+   },
+   {
+    "p50": 9.67,
+    "p95": 13.045,
+    "p99": 13.805,
+    "max": 13.805
+   },
+   {
+    "p50": 10.331,
+    "p95": 13.905,
+    "p99": 21.628,
+    "max": 21.628
+   },
+   {
+    "p50": 8.96,
+    "p95": 10.848,
+    "p99": 11.741,
+    "max": 11.741
+   }
+  ],
+  "2": [
+   {
+    "p50": 15.201,
+    "p95": 18.17,
+    "p99": 19.766,
+    "max": 19.766
+   },
+   {
+    "p50": 14.902,
+    "p95": 17.614,
+    "p99": 18.457,
+    "max": 18.457
+   },
+   {
+    "p50": 15.324,
+    "p95": 18.549,
+    "p99": 19.078,
+    "max": 19.078
+   },
+   {
+    "p50": 15.021,
+    "p95": 19.99,
+    "p99": 21.951,
+    "max": 21.951
+   },
+   {
+    "p50": 14.689,
+    "p95": 18.919,
+    "p99": 19.926,
+    "max": 19.926
+   },
+   {
+    "p50": 16.051,
+    "p95": 32.815,
+    "p99": 33.077,
+    "max": 33.077
+   },
+   {
+    "p50": 14.717,
+    "p95": 17.324,
+    "p99": 17.522,
+    "max": 17.522
+   },
+   {
+    "p50": 15.032,
+    "p95": 18.216,
+    "p99": 19.497,
+    "max": 19.497
+   },
+   {
+    "p50": 15.555,
+    "p95": 19.893,
+    "p99": 21.052,
+    "max": 21.052
+   },
+   {
+    "p50": 14.499,
+    "p95": 16.796,
+    "p99": 16.882,
+    "max": 16.882
+   },
+   {
+    "p50": 14.645,
+    "p95": 17.099,
+    "p99": 17.543,
+    "max": 17.543
+   },
+   {
+    "p50": 14.358,
+    "p95": 21.149,
+    "p99": 24.103,
+    "max": 24.103
+   },
+   {
+    "p50": 14.984,
+    "p95": 18.661,
+    "p99": 20.683,
+    "max": 20.683
+   },
+   {
+    "p50": 14.502,
+    "p95": 17.139,
+    "p99": 17.798,
+    "max": 17.798
+   },
+   {
+    "p50": 14.947,
+    "p95": 16.647,
+    "p99": 16.766,
+    "max": 16.766
+   },
+   {
+    "p50": 14.786,
+    "p95": 17.521,
+    "p99": 19.179,
+    "max": 19.179
+   },
+   {
+    "p50": 14.749,
+    "p95": 19.165,
+    "p99": 25.119,
+    "max": 25.119
+   },
+   {
+    "p50": 16.221,
+    "p95": 28.768,
+    "p99": 29.582,
+    "max": 29.582
+   },
+   {
+    "p50": 14.652,
+    "p95": 28.668,
+    "p99": 30.074,
+    "max": 30.074
+   },
+   {
+    "p50": 17.139,
+    "p95": 32.269,
+    "p99": 32.377,
+    "max": 32.377
+   },
+   {
+    "p50": 14.907,
+    "p95": 20.386,
+    "p99": 20.81,
+    "max": 20.81
+   },
+   {
+    "p50": 15.882,
+    "p95": 17.677,
+    "p99": 18.192,
+    "max": 18.192
+   },
+   {
+    "p50": 15.359,
+    "p95": 18.135,
+    "p99": 18.531,
+    "max": 18.531
+   },
+   {
+    "p50": 15.207,
+    "p95": 17.382,
+    "p99": 18.529,
+    "max": 18.529
+   }
+  ],
+  "4": [
+   {
+    "p50": 26.665,
+    "p95": 31.215,
+    "p99": 34.178,
+    "max": 34.178
+   },
+   {
+    "p50": 27.294,
+    "p95": 33.673,
+    "p99": 35.04,
+    "max": 35.04
+   },
+   {
+    "p50": 27.918,
+    "p95": 34.553,
+    "p99": 42.061,
+    "max": 42.061
+   },
+   {
+    "p50": 26.37,
+    "p95": 30.528,
+    "p99": 32.127,
+    "max": 32.127
+   },
+   {
+    "p50": 25.9,
+    "p95": 28.742,
+    "p99": 34.469,
+    "max": 34.469
+   },
+   {
+    "p50": 27.106,
+    "p95": 35.303,
+    "p99": 40.808,
+    "max": 40.808
+   },
+   {
+    "p50": 26.195,
+    "p95": 34.707,
+    "p99": 45.867,
+    "max": 45.867
+   },
+   {
+    "p50": 26.772,
+    "p95": 30.526,
+    "p99": 30.766,
+    "max": 30.766
+   },
+   {
+    "p50": 26.886,
+    "p95": 46.087,
+    "p99": 50.21,
+    "max": 50.21
+   },
+   {
+    "p50": 26.004,
+    "p95": 29.001,
+    "p99": 30.311,
+    "max": 30.311
+   },
+   {
+    "p50": 27.095,
+    "p95": 32.363,
+    "p99": 38.011,
+    "max": 38.011
+   },
+   {
+    "p50": 25.952,
+    "p95": 30.357,
+    "p99": 31.543,
+    "max": 31.543
+   },
+   {
+    "p50": 26.537,
+    "p95": 30.396,
+    "p99": 31.056,
+    "max": 31.056
+   },
+   {
+    "p50": 26.563,
+    "p95": 28.768,
+    "p99": 30.415,
+    "max": 30.415
+   },
+   {
+    "p50": 26.863,
+    "p95": 32.136,
+    "p99": 32.967,
+    "max": 32.967
+   },
+   {
+    "p50": 28.413,
+    "p95": 50.4,
+    "p99": 58.088,
+    "max": 58.088
+   },
+   {
+    "p50": 26.623,
+    "p95": 32.801,
+    "p99": 35.603,
+    "max": 35.603
+   },
+   {
+    "p50": 27.344,
+    "p95": 32.779,
+    "p99": 38.661,
+    "max": 38.661
+   },
+   {
+    "p50": 26.408,
+    "p95": 29.565,
+    "p99": 30.094,
+    "max": 30.094
+   },
+   {
+    "p50": 26.198,
+    "p95": 29.741,
+    "p99": 30.489,
+    "max": 30.489
+   },
+   {
+    "p50": 27.724,
+    "p95": 32.174,
+    "p99": 32.423,
+    "max": 32.423
+   },
+   {
+    "p50": 26.724,
+    "p95": 31.253,
+    "p99": 32.153,
+    "max": 32.153
+   },
+   {
+    "p50": 26.76,
+    "p95": 31.054,
+    "p99": 31.087,
+    "max": 31.087
+   },
+   {
+    "p50": 26.278,
+    "p95": 31.043,
+    "p99": 39.867,
+    "max": 39.867
+   }
+  ],
+  "8": [
+   {
+    "p50": 52.266,
+    "p95": 68.035,
+    "p99": 74.885,
+    "max": 74.885
+   },
+   {
+    "p50": 54.125,
+    "p95": 85.092,
+    "p99": 96.265,
+    "max": 96.265
+   },
+   {
+    "p50": 56.176,
+    "p95": 85.733,
+    "p99": 87.312,
+    "max": 87.312
+   },
+   {
+    "p50": 51.069,
+    "p95": 60.145,
+    "p99": 62.654,
+    "max": 62.654
+   },
+   {
+    "p50": 51.07,
+    "p95": 61.397,
+    "p99": 65.437,
+    "max": 65.437
+   },
+   {
+    "p50": 52.133,
+    "p95": 69.749,
+    "p99": 79.426,
+    "max": 79.426
+   },
+   {
+    "p50": 49.955,
+    "p95": 60.339,
+    "p99": 60.455,
+    "max": 60.455
+   },
+   {
+    "p50": 50.468,
+    "p95": 64.167,
+    "p99": 66.843,
+    "max": 66.843
+   },
+   {
+    "p50": 49.244,
+    "p95": 52.489,
+    "p99": 54.475,
+    "max": 54.475
+   },
+   {
+    "p50": 49.644,
+    "p95": 57.36,
+    "p99": 58.934,
+    "max": 58.934
+   },
+   {
+    "p50": 52.186,
+    "p95": 59.122,
+    "p99": 68.662,
+    "max": 68.662
+   },
+   {
+    "p50": 51.393,
+    "p95": 75.776,
+    "p99": 83.858,
+    "max": 83.858
+   },
+   {
+    "p50": 51.953,
+    "p95": 67.169,
+    "p99": 69.287,
+    "max": 69.287
+   },
+   {
+    "p50": 50.324,
+    "p95": 60.185,
+    "p99": 68.603,
+    "max": 68.603
+   },
+   {
+    "p50": 50.893,
+    "p95": 54.805,
+    "p99": 58.967,
+    "max": 58.967
+   },
+   {
+    "p50": 53.696,
+    "p95": 65.629,
+    "p99": 72.674,
+    "max": 72.674
+   },
+   {
+    "p50": 48.672,
+    "p95": 62.026,
+    "p99": 62.276,
+    "max": 62.276
+   },
+   {
+    "p50": 50.777,
+    "p95": 65.002,
+    "p99": 66.318,
+    "max": 66.318
+   },
+   {
+    "p50": 48.835,
+    "p95": 57.514,
+    "p99": 69.18,
+    "max": 69.18
+   },
+   {
+    "p50": 50.785,
+    "p95": 73.072,
+    "p99": 75.93,
+    "max": 75.93
+   },
+   {
+    "p50": 50.684,
+    "p95": 65.827,
+    "p99": 75.478,
+    "max": 75.478
+   },
+   {
+    "p50": 50.341,
+    "p95": 57.511,
+    "p99": 58.992,
+    "max": 58.992
+   },
+   {
+    "p50": 51.567,
+    "p95": 66.644,
+    "p99": 67.094,
+    "max": 67.094
+   },
+   {
+    "p50": 52.094,
+    "p95": 71.313,
+    "p99": 73.767,
+    "max": 73.767
+   }
+  ]
+ },
+ "summary": {
+  "p501": 9.019,
+  "p951": 11.911,
+  "p991": 12.272,
+  "p502": 14.984,
+  "p952": 18.549,
+  "p992": 19.766,
+  "p504": 26.724,
+  "p954": 31.253,
+  "p994": 34.178,
+  "p508": 51.069,
+  "p958": 65.002,
+  "p998": 68.662
+ },
+ "delta_vs_1step_ms": {
+  "d502": 5.965,
+  "d504": 17.705,
+  "d508": 42.05
+ }
+}

docs/experiment_round39_K16.json

@@ -0,0 +1,77 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 16,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round39_K16.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5625,
+    "collision_free_success": 0.5625,
+    "time_ok_1p3": 0.296875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.6243183612823486,
+    "time_mean_s": 2.6024999618530273,
+    "brakes_mean": 13.7734375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2799961566925049,
+    "err_fail_med": 195.448486328125,
+    "err_success_med": 0.15947841107845306,
+    "dls_share": 0.640625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 1.5750000476837158,
+    "reverse_mean": 1.6161863803863525
+  },
+  "simulation_wall_s": 573.4911701000092
+}

docs/experiment_round39_lenw0.1.json

@@ -0,0 +1,75 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round39_lenw0.1.json",
+    "trace": null,
+    "select": "ep_len",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.1,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.51171875,
+    "collision_free_success": 0.51171875,
+    "time_ok_1p3": 0.26953125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.621551036834717,
+    "time_mean_s": 2.4761831760406494,
+    "brakes_mean": 19.48828125,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2874715030193329,
+    "err_fail_med": 208.0048065185547,
+    "err_success_med": 0.24376358091831207,
+    "dls_share": 0.6171875,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118
+  },
+  "simulation_wall_s": 333.6044210000109
+}

docs/experiment_round39_lenw0.3.json

@@ -0,0 +1,75 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round39_lenw0.3.json",
+    "trace": null,
+    "select": "ep_len",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.3,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.51953125,
+    "collision_free_success": 0.51953125,
+    "time_ok_1p3": 0.28515625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.6547701358795166,
+    "time_mean_s": 2.4798495769500732,
+    "brakes_mean": 20.00390625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2706023156642914,
+    "err_fail_med": 208.01719665527344,
+    "err_success_med": 0.2188246101140976,
+    "dls_share": 0.62109375,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118
+  },
+  "simulation_wall_s": 340.16115129999525
+}

docs/experiment_round39_lenw1.0.json

@@ -0,0 +1,75 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round39_lenw1.0.json",
+    "trace": null,
+    "select": "ep_len",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 1.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.52734375,
+    "collision_free_success": 0.52734375,
+    "time_ok_1p3": 0.2890625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.6718690395355225,
+    "time_mean_s": 2.463407516479492,
+    "brakes_mean": 19.3515625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2807892858982086,
+    "err_fail_med": 205.99014282226562,
+    "err_success_med": 0.23181582987308502,
+    "dls_share": 0.62890625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118
+  },
+  "simulation_wall_s": 368.1434382000007
+}

docs/experiment_round39_margin_audit.json

@@ -0,0 +1,32 @@
+{
+ "n_tasks": 256,
+ "n_plans": 195,
+ "cap_mm": 0.5,
+ "run_points": 169216,
+ "gap_hist_of_run_points": {
+  "<0.02": 0.0025056733284145594,
+  "<0.05": 0.0043317414820194244,
+  "<0.10": 0.008356183767318726,
+  "<0.20": 0.019064391031861305,
+  "<0.50": 0.055952154099941254
+ },
+ "at_cap_share": 0.9097898602485657,
+ "min_gap_per_task_p50": 0.02924060821533203,
+ "plan_kind": {
+  "full": 0.7935897707939148,
+  "partial": 0.10769230872392654,
+  "rejected": 0.09871795028448105
+ },
+ "per_task": {
+  "plans_p50": 195.0,
+  "committed_steps_share": 0.9871794581413269,
+  "rejected_share_by_group": {
+   "arrived": 0.008826094679534435,
+   "failed": 0.21994824707508087
+  }
+ },
+ "rejected_step_share_by_group": {
+  "arrived": 0.01765218935906887,
+  "failed": 0.43989649415016174
+ }
+}

docs/experiment_round39_reverse_offline.json

@@ -0,0 +1,16 @@
+[
+ {
+  "file": "round33_cap.pt",
+  "reverse_p50": 0.2474,
+  "reverse_p50_arrived": 0.265,
+  "reverse_p50_failed": 0.2329,
+  "nact_p50": 168.0
+ },
+ {
+  "file": "round33_base.pt",
+  "reverse_p50": 0.257,
+  "reverse_p50_arrived": 0.2753,
+  "reverse_p50_failed": 0.2402,
+  "nact_p50": 152.0
+ }
+]

docs/experiment_round39_travel_decomp.json

@@ -0,0 +1,28 @@
+{
+ "n": 256,
+ "travel_ratio_median": 2.425084623442128,
+ "travel_ratio_population": "arrived & 有老师参考（与 eval 的 okr 同口径）",
+ "reverse_rate": {
+  "p25": 0.15732546150684357,
+  "p50": 0.23536035418510437,
+  "p75": 0.3025793731212616
+ },
+ "net_over_travel_p50": 0.4447892699092449,
+ "excursion_over_travel_p50": 0.5067885044066773,
+ "still_rate_p50": 0.028846153989434242,
+ "by_group": {
+  "arrived": {
+   "n": 147,
+   "reverse_p50": 0.26143792271614075,
+   "net_over_travel_p50": 0.47818613856136644,
+   "excursion_p50": 0.5500977881151917
+  },
+  "failed": {
+   "n": 109,
+   "reverse_p50": 0.14653243124485016,
+   "net_over_travel_p50": 0.3515762667187255,
+   "excursion_p50": 0.39037085057225707
+  }
+ },
+ "note": "net/行程 越接近 1 越\"单向\"；越小越来回抖。excursion/行程 = 各轴走过最大幅度的和 / 总行程，绕远时它接近 1、抖动时明显小于 1"
+}

docs/experiment_round39_try1.json

@@ -0,0 +1,77 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 1,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round39_try1.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.40234375,
+    "collision_free_success": 0.40234375,
+    "time_ok_1p3": 0.22265625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.5106539726257324,
+    "time_mean_s": 2.253203868865967,
+    "brakes_mean": 60.44140625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.4367985725402832,
+    "err_fail_med": 243.9196319580078,
+    "err_success_med": 0.21751396358013153,
+    "dls_share": 0.53125,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000038146972656,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.18634259700775146,
+    "reverse_mean": 0.19426266849040985
+  },
+  "simulation_wall_s": 111.52556069999991
+}

docs/experiment_round39_try3.json

@@ -0,0 +1,77 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 3,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round39_try3.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.45703125,
+    "collision_free_success": 0.45703125,
+    "time_ok_1p3": 0.23828125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.529658794403076,
+    "time_mean_s": 2.3210256099700928,
+    "brakes_mean": 37.75390625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.3564552068710327,
+    "err_fail_med": 220.24862670898438,
+    "err_success_med": 0.23727791011333466,
+    "dls_share": 0.56640625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.22442243993282318,
+    "reverse_mean": 0.23038837313652039
+  },
+  "simulation_wall_s": 170.3933407000004
+}

docs/experiment_round39_z15.json

@@ -0,0 +1,79 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round39_z15.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 1,
+    "z_scale": 1.5,
+    "len_w": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.49609375,
+    "collision_free_success": 0.49609375,
+    "time_ok_1p3": 0.18359375,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 3.52425217628479,
+    "time_mean_s": 2.805511713027954,
+    "brakes_mean": 10.51171875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.27530819177627563,
+    "err_fail_med": 105.35015869140625,
+    "err_success_med": 0.23657508194446564,
+    "dls_share": 0.62890625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.32083332538604736,
+    "reverse_mean": 0.3213125765323639,
+    "reverse_moving_p50": 0.35158151388168335,
+    "moving_share_p50": 0.9866666793823242
+  },
+  "simulation_wall_s": 311.9796243000019
+}

docs/experiment_round40_K16cb1.json

@@ -0,0 +1,82 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 16,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round40_K16cb1.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5625,
+    "collision_free_success": 0.5625,
+    "time_ok_1p3": 0.296875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.6243183612823486,
+    "time_mean_s": 2.6024999618530273,
+    "brakes_mean": 13.7734375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2799961566925049,
+    "err_fail_med": 195.448486328125,
+    "err_success_med": 0.15947841107845306,
+    "dls_share": 0.640625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.26249998807907104,
+    "reverse_mean": 0.2693643867969513,
+    "reverse_moving_p50": 0.3194444477558136,
+    "moving_share_p50": 0.9869565367698669,
+    "brake_rate_mean": 0.07063301652669907,
+    "brake_rate_p50": 0.0,
+    "n_plans_mean": 195.0
+  },
+  "simulation_wall_s": 469.2955887999997
+}

docs/experiment_round40_k_cost.json

@@ -0,0 +1,226 @@
+{
+ "device": "NVIDIA GeForce RTX 3090",
+ "B": 1,
+ "inc_cap": 1,
+ "infer": {
+  "8": [
+   {
+    "p50": 9.285,
+    "p95": 10.62
+   },
+   {
+    "p50": 9.227,
+    "p95": 15.144
+   },
+   {
+    "p50": 10.015,
+    "p95": 18.492
+   },
+   {
+    "p50": 10.114,
+    "p95": 15.477
+   }
+  ],
+  "16": [
+   {
+    "p50": 9.281,
+    "p95": 12.2
+   },
+   {
+    "p50": 9.321,
+    "p95": 15.859
+   },
+   {
+    "p50": 11.673,
+    "p95": 17.75
+   },
+   {
+    "p50": 9.22,
+    "p95": 11.764
+   }
+  ]
+ },
+ "plan": {
+  "8/cb0": [
+   {
+    "p50": 134.514,
+    "p95": 146.198,
+    "max": 151.847
+   },
+   {
+    "p50": 147.029,
+    "p95": 180.064,
+    "max": 191.364
+   },
+   {
+    "p50": 133.562,
+    "p95": 145.323,
+    "max": 150.765
+   },
+   {
+    "p50": 141.62,
+    "p95": 163.992,
+    "max": 194.319
+   },
+   {
+    "p50": 146.427,
+    "p95": 186.324,
+    "max": 189.776
+   },
+   {
+    "p50": 144.785,
+    "p95": 164.472,
+    "max": 168.658
+   },
+   {
+    "p50": 144.466,
+    "p95": 204.952,
+    "max": 227.804
+   },
+   {
+    "p50": 151.671,
+    "p95": 189.823,
+    "max": 241.949
+   }
+  ],
+  "16/cb0": [
+   {
+    "p50": 133.572,
+    "p95": 153.984,
+    "max": 156.276
+   },
+   {
+    "p50": 144.001,
+    "p95": 170.74,
+    "max": 185.805
+   },
+   {
+    "p50": 139.492,
+    "p95": 162.124,
+    "max": 165.545
+   },
+   {
+    "p50": 139.614,
+    "p95": 152.373,
+    "max": 153.8
+   },
+   {
+    "p50": 137.178,
+    "p95": 158.119,
+    "max": 161.382
+   },
+   {
+    "p50": 136.955,
+    "p95": 175.71,
+    "max": 186.645
+   },
+   {
+    "p50": 143.252,
+    "p95": 166.76,
+    "max": 182.599
+   },
+   {
+    "p50": 149.9,
+    "p95": 200.22,
+    "max": 201.094
+   }
+  ],
+  "8/cb1": [
+   {
+    "p50": 151.167,
+    "p95": 164.805,
+    "max": 187.036
+   },
+   {
+    "p50": 148.91,
+    "p95": 168.584,
+    "max": 177.316
+   },
+   {
+    "p50": 162.24,
+    "p95": 204.403,
+    "max": 222.327
+   },
+   {
+    "p50": 158.656,
+    "p95": 197.649,
+    "max": 224.336
+   },
+   {
+    "p50": 149.629,
+    "p95": 176.731,
+    "max": 187.429
+   },
+   {
+    "p50": 149.44,
+    "p95": 200.478,
+    "max": 243.552
+   },
+   {
+    "p50": 150.726,
+    "p95": 200.097,
+    "max": 242.956
+   },
+   {
+    "p50": 146.359,
+    "p95": 165.303,
+    "max": 165.497
+   }
+  ],
+  "16/cb1": [
+   {
+    "p50": 156.341,
+    "p95": 180.399,
+    "max": 184.267
+   },
+   {
+    "p50": 169.067,
+    "p95": 206.577,
+    "max": 218.282
+   },
+   {
+    "p50": 169.66,
+    "p95": 190.207,
+    "max": 202.072
+   },
+   {
+    "p50": 163.572,
+    "p95": 185.455,
+    "max": 192.382
+   },
+   {
+    "p50": 172.023,
+    "p95": 204.316,
+    "max": 225.098
+   },
+   {
+    "p50": 166.825,
+    "p95": 201.437,
+    "max": 236.891
+   },
+   {
+    "p50": 171.442,
+    "p95": 207.533,
+    "max": 226.472
+   },
+   {
+    "p50": 159.815,
+    "p95": 198.534,
+    "max": 207.788
+   }
+  ]
+ },
+ "summary": {
+  "infer_p50_K8": 10.015,
+  "infer_p50_K16": 9.321,
+  "plan_p50_8/cb0": 144.785,
+  "plan_p50_16/cb0": 139.614,
+  "plan_p50_8/cb1": 150.726,
+  "plan_p50_16/cb1": 169.067,
+  "plan_p95_8/cb0": 180.064,
+  "plan_p95_16/cb0": 166.76,
+  "plan_p95_8/cb1": 197.649,
+  "plan_p95_16/cb1": 201.437
+ },
+ "note": "K16/try3 那批臂是串行认证(cert_batch=0)，最坏要 K 次 gaps 调用；部署口径应是批量认证=1 次调用，所以两种口径都要量"
+}

docs/experiment_round40_safetyoff_steps1.json

@@ -0,0 +1,82 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round40_safetyoff_steps1.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 0,
+    "partial": 1,
+    "inc_cap": 0,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.77734375,
+    "collision_free_success": 0.359375,
+    "time_ok_1p3": 0.53515625,
+    "collision_rate": 0.5859375,
+    "collision_steps_mean": 58.33984375,
+    "travel_ratio_median": 2.375220775604248,
+    "time_mean_s": 2.3642210960388184,
+    "brakes_mean": 0.0,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.0402195081114769,
+    "err_fail_med": 50.388858795166016,
+    "err_success_med": 0.1961677223443985,
+    "dls_share": 0.84765625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.6600004434585571,
+    "speed_ratio_p95": 1.660003662109375,
+    "speed_ratio_max": 1.660003662109375,
+    "over_vmax_rate": 1.0,
+    "accel_ratio_p95": 1.5426017045974731,
+    "accel_ratio_max": 1.6600004434585571,
+    "reverse_p50": 0.2846153974533081,
+    "reverse_mean": 0.2833859324455261,
+    "reverse_moving_p50": 0.2846153974533081,
+    "moving_share_p50": 1.0,
+    "brake_rate_mean": 0.0,
+    "brake_rate_p50": 0.0,
+    "n_plans_mean": 195.0
+  },
+  "simulation_wall_s": 52.50466950000555
+}

docs/experiment_round40_safetyoff_steps2.json

@@ -0,0 +1,82 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round40_safetyoff_steps2.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 0,
+    "partial": 1,
+    "inc_cap": 0,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 2,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.81640625,
+    "collision_free_success": 0.375,
+    "time_ok_1p3": 0.68359375,
+    "collision_rate": 0.55859375,
+    "collision_steps_mean": 45.07421875,
+    "travel_ratio_median": 1.5936355590820312,
+    "time_mean_s": 2.019808530807495,
+    "brakes_mean": 0.0,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.003008240135386586,
+    "err_fail_med": 48.19245529174805,
+    "err_success_med": 0.2415877878665924,
+    "dls_share": 0.8515625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.6600004434585571,
+    "speed_ratio_p95": 1.660003662109375,
+    "speed_ratio_max": 1.660003662109375,
+    "over_vmax_rate": 0.9921875,
+    "accel_ratio_p95": 1.010289192199707,
+    "accel_ratio_max": 1.4423871040344238,
+    "reverse_p50": 0.08469945192337036,
+    "reverse_mean": 0.09047156572341919,
+    "reverse_moving_p50": 0.08469945192337036,
+    "moving_share_p50": 1.0,
+    "brake_rate_mean": 0.0,
+    "brake_rate_p50": 0.0,
+    "n_plans_mean": 195.0
+  },
+  "simulation_wall_s": 58.29419349999807
+}

docs/experiment_round40_waste0.1.json

@@ -0,0 +1,82 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round40_waste0.1.json",
+    "trace": null,
+    "select": "ep_waste",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.1,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.515625,
+    "collision_free_success": 0.515625,
+    "time_ok_1p3": 0.26953125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.669708251953125,
+    "time_mean_s": 2.452272653579712,
+    "brakes_mean": 19.39453125,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.28827616572380066,
+    "err_fail_med": 201.7505645751953,
+    "err_success_med": 0.20423255860805511,
+    "dls_share": 0.609375,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.2527472674846649,
+    "reverse_mean": 0.262615442276001,
+    "reverse_moving_p50": 0.31912145018577576,
+    "moving_share_p50": 0.9599999785423279,
+    "brake_rate_mean": 0.09945914149284363,
+    "brake_rate_p50": 0.0051282052882015705,
+    "n_plans_mean": 195.0
+  },
+  "simulation_wall_s": 308.12369770000805
+}

docs/experiment_round40_waste0.3.json

@@ -0,0 +1,82 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round40_waste0.3.json",
+    "trace": null,
+    "select": "ep_waste",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.3,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.515625,
+    "collision_free_success": 0.515625,
+    "time_ok_1p3": 0.265625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.5349059104919434,
+    "time_mean_s": 2.476060628890991,
+    "brakes_mean": 19.62890625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.31392601132392883,
+    "err_fail_med": 200.5723876953125,
+    "err_success_med": 0.16847863793373108,
+    "dls_share": 0.609375,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.25111111998558044,
+    "reverse_mean": 0.26184794306755066,
+    "reverse_moving_p50": 0.3268229067325592,
+    "moving_share_p50": 0.9636363387107849,
+    "brake_rate_mean": 0.10066106170415878,
+    "brake_rate_p50": 0.0051282052882015705,
+    "n_plans_mean": 195.0
+  },
+  "simulation_wall_s": 299.3823047999904
+}

docs/experiment_round40_waste1.0.json

@@ -0,0 +1,82 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round40_waste1.0.json",
+    "trace": null,
+    "select": "ep_waste",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 1.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.48828125,
+    "collision_free_success": 0.48828125,
+    "time_ok_1p3": 0.23828125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.708263874053955,
+    "time_mean_s": 2.5168001651763916,
+    "brakes_mean": 20.02734375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.3584502637386322,
+    "err_fail_med": 190.41900634765625,
+    "err_success_med": 0.33570292592048645,
+    "dls_share": 0.60546875,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.24899598956108093,
+    "reverse_mean": 0.2592279314994812,
+    "reverse_moving_p50": 0.3214285671710968,
+    "moving_share_p50": 0.9587156176567078,
+    "brake_rate_mean": 0.10270433127880096,
+    "brake_rate_p50": 0.010256410576403141,
+    "n_plans_mean": 195.0
+  },
+  "simulation_wall_s": 297.2535624000011
+}

docs/experiment_round42_K16cb1_dump.json

@@ -0,0 +1,107 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 16,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round42_K16cb1_rows.json",
+    "out": "docs/experiment_round42_K16cb1_dump.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5625,
+    "collision_free_success": 0.5625,
+    "time_ok_1p3": 0.296875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.6243183612823486,
+    "time_mean_s": 2.6024999618530273,
+    "brakes_mean": 13.7734375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2799961566925049,
+    "err_fail_med": 195.448486328125,
+    "err_success_med": 0.15947841107845306,
+    "dls_share": 0.640625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.26249998807907104,
+    "reverse_mean": 0.2693643867969513,
+    "reverse_moving_p50": 0.3194444477558136,
+    "moving_share_p50": 0.9869565367698669,
+    "brake_rate_mean": 0.07063301652669907,
+    "brake_rate_p50": 0.0,
+    "brake_rate_p90": 0.2589743733406067,
+    "stuck_share": 0.15234375,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5925925970077515,
+    "by_cat": {
+      "A": [
+        128,
+        100
+      ],
+      "B": [
+        74,
+        24
+      ],
+      "C": [
+        32,
+        12
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        8
+      ]
+    }
+  },
+  "simulation_wall_s": 613.4877918999991
+}

docs/experiment_round42_K8cb1_dump.json

@@ -0,0 +1,107 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round42_K8cb1_rows.json",
+    "out": "docs/experiment_round42_K8cb1_dump.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5234375,
+    "collision_free_success": 0.5234375,
+    "time_ok_1p3": 0.265625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.615795850753784,
+    "time_mean_s": 2.486567258834839,
+    "brakes_mean": 19.5546875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.27920001745224,
+    "err_fail_med": 204.71360778808594,
+    "err_success_med": 0.21751396358013153,
+    "dls_share": 0.61328125,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.24741201102733612,
+    "reverse_mean": 0.25991496443748474,
+    "reverse_moving_p50": 0.32581019401550293,
+    "moving_share_p50": 0.9587156176567078,
+    "brake_rate_mean": 0.10028044879436493,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.3153846263885498,
+    "stuck_share": 0.234375,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5514403581619263,
+    "by_cat": {
+      "A": [
+        128,
+        96
+      ],
+      "B": [
+        74,
+        20
+      ],
+      "C": [
+        32,
+        11
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    }
+  },
+  "simulation_wall_s": 397.44336470001144
+}

docs/experiment_round43_control_strat.json

@@ -0,0 +1,107 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round43_control_strat.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 0,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5234375,
+    "collision_free_success": 0.5234375,
+    "time_ok_1p3": 0.265625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.615795850753784,
+    "time_mean_s": 2.486567258834839,
+    "brakes_mean": 19.5546875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.27920001745224,
+    "err_fail_med": 204.71360778808594,
+    "err_success_med": 0.21751396358013153,
+    "dls_share": 0.61328125,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.24741201102733612,
+    "reverse_mean": 0.25991496443748474,
+    "reverse_moving_p50": 0.32581019401550293,
+    "moving_share_p50": 0.9587156176567078,
+    "brake_rate_mean": 0.10028044879436493,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.3153846263885498,
+    "stuck_share": 0.234375,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5514403581619263,
+    "by_cat": {
+      "A": [
+        128,
+        96
+      ],
+      "B": [
+        74,
+        20
+      ],
+      "C": [
+        32,
+        11
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    }
+  },
+  "simulation_wall_s": 353.7248780000082
+}

docs/experiment_round44_tieprobe64.json

@@ -0,0 +1,111 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 64,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round44_tieprobe64_rows.json",
+    "out": "docs/experiment_round44_tieprobe64.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 1,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "6185a312dbae3bf5b361c4b9ff3aa11b7329d9723cff2be9b9157563d6b6ec1c",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 64,
+    "n_ref": 64,
+    "success": 0.40625,
+    "collision_free_success": 0.40625,
+    "time_ok_1p3": 0.234375,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.358931303024292,
+    "time_mean_s": 2.3323075771331787,
+    "brakes_mean": 18.71875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.33309799432754517,
+    "err_fail_med": 229.02410888671875,
+    "err_success_med": 0.22094391286373138,
+    "dls_share": 0.515625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000091791152954,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.25150150060653687,
+    "reverse_mean": 0.25861242413520813,
+    "reverse_moving_p50": 0.318452388048172,
+    "moving_share_p50": 0.9090909361839294,
+    "brake_rate_mean": 0.10696429014205933,
+    "brake_rate_p50": 0.017142856493592262,
+    "brake_rate_p90": 0.33142855763435364,
+    "stuck_share": 0.21875,
+    "n_plans_mean": 175.0,
+    "success_reachable": 0.4262295067310333,
+    "by_cat": {
+      "A": [
+        28,
+        14
+      ],
+      "B": [
+        21,
+        6
+      ],
+      "C": [
+        9,
+        4
+      ],
+      "D": [
+        3,
+        0
+      ],
+      "Dr": [
+        3,
+        2
+      ]
+    },
+    "tie_opp_rate_mean": 0.7484821081161499,
+    "tie_changed_rate_mean": 0.07437500357627869,
+    "tie_opp_task_share": 1.0
+  },
+  "simulation_wall_s": 78.19308080000337
+}

docs/experiment_round45_tie0.json

@@ -0,0 +1,108 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round45_tie0_rows.json",
+    "out": "docs/experiment_round45_tie0.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5234375,
+    "collision_free_success": 0.5234375,
+    "time_ok_1p3": 0.265625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.615795850753784,
+    "time_mean_s": 2.486567258834839,
+    "brakes_mean": 19.5546875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.27920001745224,
+    "err_fail_med": 204.71360778808594,
+    "err_success_med": 0.21751396358013153,
+    "dls_share": 0.61328125,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.24741201102733612,
+    "reverse_mean": 0.25991496443748474,
+    "reverse_moving_p50": 0.32581019401550293,
+    "moving_share_p50": 0.9587156176567078,
+    "brake_rate_mean": 0.10028044879436493,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.3153846263885498,
+    "stuck_share": 0.234375,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5514403581619263,
+    "by_cat": {
+      "A": [
+        128,
+        96
+      ],
+      "B": [
+        74,
+        20
+      ],
+      "C": [
+        32,
+        11
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    }
+  },
+  "simulation_wall_s": 317.56223549999413
+}

docs/experiment_round45_tie1.json

@@ -0,0 +1,111 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round45_tie1_rows.json",
+    "out": "docs/experiment_round45_tie1.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 1,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.50390625,
+    "collision_free_success": 0.50390625,
+    "time_ok_1p3": 0.27734375,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.53989315032959,
+    "time_mean_s": 2.3586044311523438,
+    "brakes_mean": 17.7890625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.27920001745224,
+    "err_fail_med": 213.06358337402344,
+    "err_success_med": 0.22968679666519165,
+    "dls_share": 0.62109375,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.25315314531326294,
+    "reverse_mean": 0.26444095373153687,
+    "reverse_moving_p50": 0.3198198080062866,
+    "moving_share_p50": 0.9772727489471436,
+    "brake_rate_mean": 0.09122596681118011,
+    "brake_rate_p50": 0.0,
+    "brake_rate_p90": 0.3282051384449005,
+    "stuck_share": 0.203125,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5308641791343689,
+    "by_cat": {
+      "A": [
+        128,
+        91
+      ],
+      "B": [
+        74,
+        19
+      ],
+      "C": [
+        32,
+        12
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    },
+    "tie_opp_rate_mean": 0.7898838520050049,
+    "tie_changed_rate_mean": 0.055829327553510666,
+    "tie_opp_task_share": 1.0
+  },
+  "simulation_wall_s": 283.46215640001174
+}

docs/experiment_round46_ep16_smoke.json

@@ -0,0 +1,96 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 16,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round46_ep16_smoke.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "7b7c219c70d15b280682fe0626192b267f0e4dfd843b1e70fb279e148eb6eaa3",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 16,
+    "n_ref": 16,
+    "success": 0.375,
+    "collision_free_success": 0.375,
+    "time_ok_1p3": 0.25,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.426271438598633,
+    "time_mean_s": 2.680000066757202,
+    "brakes_mean": 24.9375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.43800246715545654,
+    "err_fail_med": 231.5127410888672,
+    "err_success_med": 0.2731371223926544,
+    "dls_share": 0.5,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000053644180298,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000016689300537,
+    "reverse_p50": 0.22442243993282318,
+    "reverse_mean": 0.2425607144832611,
+    "reverse_moving_p50": 0.30572596192359924,
+    "moving_share_p50": 0.7567567825317383,
+    "brake_rate_mean": 0.14249999821186066,
+    "brake_rate_p50": 0.07428571581840515,
+    "brake_rate_p90": 0.3571428656578064,
+    "stuck_share": 0.3125,
+    "n_plans_mean": 175.0,
+    "success_reachable": 0.375,
+    "by_cat": {
+      "A": [
+        8,
+        4
+      ],
+      "B": [
+        8,
+        2
+      ]
+    }
+  },
+  "simulation_wall_s": 35.48499080000329
+}

docs/experiment_round46_pre16_smoke.json

@@ -0,0 +1,96 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 16,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": null,
+    "out": "docs/experiment_round46_pre16_smoke.json",
+    "trace": null,
+    "select": "pre",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "7b7c219c70d15b280682fe0626192b267f0e4dfd843b1e70fb279e148eb6eaa3",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 16,
+    "n_ref": 16,
+    "success": 0.375,
+    "collision_free_success": 0.375,
+    "time_ok_1p3": 0.25,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.6735622882843018,
+    "time_mean_s": 2.763333320617676,
+    "brakes_mean": 21.9375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.4253818392753601,
+    "err_fail_med": 223.724365234375,
+    "err_success_med": 0.7986489534378052,
+    "dls_share": 0.5625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000053644180298,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000016689300537,
+    "reverse_p50": 0.23239436745643616,
+    "reverse_mean": 0.22883012890815735,
+    "reverse_moving_p50": 0.29054054617881775,
+    "moving_share_p50": 0.9583333134651184,
+    "brake_rate_mean": 0.12535713613033295,
+    "brake_rate_p50": 0.0057142856530845165,
+    "brake_rate_p90": 0.3257142901420593,
+    "stuck_share": 0.3125,
+    "n_plans_mean": 175.0,
+    "success_reachable": 0.375,
+    "by_cat": {
+      "A": [
+        8,
+        4
+      ],
+      "B": [
+        8,
+        2
+      ]
+    }
+  },
+  "simulation_wall_s": 31.1417752999987
+}

docs/experiment_round46_pre256.json

@@ -0,0 +1,108 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round46_pre256_rows.json",
+    "out": "docs/experiment_round46_pre256.json",
+    "trace": null,
+    "select": "pre",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.54296875,
+    "collision_free_success": 0.54296875,
+    "time_ok_1p3": 0.33203125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.4448752403259277,
+    "time_mean_s": 2.342589855194092,
+    "brakes_mean": 19.83984375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.17014972865581512,
+    "err_fail_med": 185.13986206054688,
+    "err_success_med": 0.4150978922843933,
+    "dls_share": 0.63671875,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.22891566157341003,
+    "reverse_mean": 0.23401273787021637,
+    "reverse_moving_p50": 0.28184282779693604,
+    "moving_share_p50": 0.9583333134651184,
+    "brake_rate_mean": 0.10174279659986496,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.35384616255760193,
+    "stuck_share": 0.24609375,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5720164775848389,
+    "by_cat": {
+      "A": [
+        128,
+        95
+      ],
+      "B": [
+        74,
+        22
+      ],
+      "C": [
+        32,
+        14
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        8
+      ]
+    }
+  },
+  "simulation_wall_s": 319.2378991000005
+}

docs/experiment_round47_ep_s1.json

@@ -0,0 +1,108 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round47_ep_s1_rows.json",
+    "out": "docs/experiment_round47_ep_s1.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 1,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.53125,
+    "collision_free_success": 0.53125,
+    "time_ok_1p3": 0.296875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.636941432952881,
+    "time_mean_s": 2.4600000381469727,
+    "brakes_mean": 19.265625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2750342786312103,
+    "err_fail_med": 214.6193084716797,
+    "err_success_med": 0.22276456654071808,
+    "dls_share": 0.62890625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.25339367985725403,
+    "reverse_mean": 0.2575649917125702,
+    "reverse_moving_p50": 0.3117443919181824,
+    "moving_share_p50": 0.9636363387107849,
+    "brake_rate_mean": 0.09879808127880096,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.3384615480899811,
+    "stuck_share": 0.25390625,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5596708059310913,
+    "by_cat": {
+      "A": [
+        128,
+        96
+      ],
+      "B": [
+        74,
+        21
+      ],
+      "C": [
+        32,
+        12
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    }
+  },
+  "simulation_wall_s": 315.10028919999604
+}

docs/experiment_round47_pre_s1.json

@@ -0,0 +1,108 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round47_pre_s1_rows.json",
+    "out": "docs/experiment_round47_pre_s1.json",
+    "trace": null,
+    "select": "pre",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 1,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.54296875,
+    "collision_free_success": 0.54296875,
+    "time_ok_1p3": 0.33984375,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.4660637378692627,
+    "time_mean_s": 2.4476258754730225,
+    "brakes_mean": 19.703125,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.20109058916568756,
+    "err_fail_med": 187.68463134765625,
+    "err_success_med": 0.3424937129020691,
+    "dls_share": 0.6484375,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.22884614765644073,
+    "reverse_mean": 0.23487445712089539,
+    "reverse_moving_p50": 0.2777777910232544,
+    "moving_share_p50": 0.9679999947547913,
+    "brake_rate_mean": 0.10104167461395264,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.35641026496887207,
+    "stuck_share": 0.2421875,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5720164775848389,
+    "by_cat": {
+      "A": [
+        128,
+        95
+      ],
+      "B": [
+        74,
+        23
+      ],
+      "C": [
+        32,
+        14
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    }
+  },
+  "simulation_wall_s": 314.3736673000094
+}

docs/experiment_round48_pre_h4.json

@@ -0,0 +1,109 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round48_pre_h4_rows.json",
+    "out": "docs/experiment_round48_pre_h4.json",
+    "trace": null,
+    "select": "pre",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "pre_h": 4,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5390625,
+    "collision_free_success": 0.5390625,
+    "time_ok_1p3": 0.3046875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.5618672370910645,
+    "time_mean_s": 2.4820291996002197,
+    "brakes_mean": 20.40625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.23497404158115387,
+    "err_fail_med": 199.24868774414062,
+    "err_success_med": 0.33005011081695557,
+    "dls_share": 0.62890625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.24242424964904785,
+    "reverse_mean": 0.24662044644355774,
+    "reverse_moving_p50": 0.30327868461608887,
+    "moving_share_p50": 0.9583333134651184,
+    "brake_rate_mean": 0.1046474277973175,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.3589743673801422,
+    "stuck_share": 0.24609375,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5679012537002563,
+    "by_cat": {
+      "A": [
+        128,
+        96
+      ],
+      "B": [
+        74,
+        22
+      ],
+      "C": [
+        32,
+        13
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    }
+  },
+  "simulation_wall_s": 349.25640099999146
+}

docs/experiment_round48_pre_h8.json

@@ -0,0 +1,109 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round48_pre_h8_rows.json",
+    "out": "docs/experiment_round48_pre_h8.json",
+    "trace": null,
+    "select": "pre",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "pre_h": 8,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.51953125,
+    "collision_free_success": 0.51953125,
+    "time_ok_1p3": 0.28125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.6749093532562256,
+    "time_mean_s": 2.445413589477539,
+    "brakes_mean": 20.26953125,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2660874128341675,
+    "err_fail_med": 201.30308532714844,
+    "err_success_med": 0.22553904354572296,
+    "dls_share": 0.6328125,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.25132274627685547,
+    "reverse_mean": 0.25476565957069397,
+    "reverse_moving_p50": 0.31292515993118286,
+    "moving_share_p50": 0.9629629850387573,
+    "brake_rate_mean": 0.10394632816314697,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.34871795773506165,
+    "stuck_share": 0.2578125,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.547325074672699,
+    "by_cat": {
+      "A": [
+        128,
+        92
+      ],
+      "B": [
+        74,
+        20
+      ],
+      "C": [
+        32,
+        14
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    }
+  },
+  "simulation_wall_s": 338.30686029999924
+}

docs/experiment_round49_gate20.json

@@ -0,0 +1,110 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round49_gate20_rows.json",
+    "out": "docs/experiment_round49_gate20.json",
+    "trace": null,
+    "select": "pre",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "pre_h": 0,
+    "pre_near": 20.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.53515625,
+    "collision_free_success": 0.53515625,
+    "time_ok_1p3": 0.328125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.5145769119262695,
+    "time_mean_s": 2.3382482528686523,
+    "brakes_mean": 19.63671875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.17014972865581512,
+    "err_fail_med": 182.09060668945312,
+    "err_success_med": 0.2248138189315796,
+    "dls_share": 0.63671875,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.24203822016716003,
+    "reverse_mean": 0.24321714043617249,
+    "reverse_moving_p50": 0.2936508059501648,
+    "moving_share_p50": 0.9595959782600403,
+    "brake_rate_mean": 0.10070113092660904,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.35384616255760193,
+    "stuck_share": 0.2421875,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5637860298156738,
+    "by_cat": {
+      "A": [
+        128,
+        94
+      ],
+      "B": [
+        74,
+        21
+      ],
+      "C": [
+        32,
+        14
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        8
+      ]
+    }
+  },
+  "simulation_wall_s": 285.5599780000048
+}

docs/experiment_round49_gate5.json

@@ -0,0 +1,110 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round49_gate5_rows.json",
+    "out": "docs/experiment_round49_gate5.json",
+    "trace": null,
+    "select": "pre",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "pre_h": 0,
+    "pre_near": 5.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.546875,
+    "collision_free_success": 0.546875,
+    "time_ok_1p3": 0.33203125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.4217512607574463,
+    "time_mean_s": 2.3379998207092285,
+    "brakes_mean": 19.83984375,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.17014972865581512,
+    "err_fail_med": 185.13986206054688,
+    "err_success_med": 0.3768002390861511,
+    "dls_share": 0.63671875,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.23239436745643616,
+    "reverse_mean": 0.23485617339611053,
+    "reverse_moving_p50": 0.28134557604789734,
+    "moving_share_p50": 0.9583333134651184,
+    "brake_rate_mean": 0.10174278169870377,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.35384616255760193,
+    "stuck_share": 0.24609375,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5761317014694214,
+    "by_cat": {
+      "A": [
+        128,
+        96
+      ],
+      "B": [
+        74,
+        22
+      ],
+      "C": [
+        32,
+        14
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        8
+      ]
+    }
+  },
+  "simulation_wall_s": 332.8420308999921
+}

docs/experiment_round50_gate20_s1.json

@@ -0,0 +1,110 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round50_gate20_s1_rows.json",
+    "out": "docs/experiment_round50_gate20_s1.json",
+    "trace": null,
+    "select": "pre",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "pre_h": 0,
+    "pre_near": 20.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 1,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.53125,
+    "collision_free_success": 0.53125,
+    "time_ok_1p3": 0.30078125,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.6468124389648438,
+    "time_mean_s": 2.462352752685547,
+    "brakes_mean": 19.71875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.20109058916568756,
+    "err_fail_med": 177.47691345214844,
+    "err_success_med": 0.22864246368408203,
+    "dls_share": 0.6484375,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.2363184094429016,
+    "reverse_mean": 0.24333104491233826,
+    "reverse_moving_p50": 0.28763440251350403,
+    "moving_share_p50": 0.9736841917037964,
+    "brake_rate_mean": 0.10112179815769196,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.35641026496887207,
+    "stuck_share": 0.24609375,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5596708059310913,
+    "by_cat": {
+      "A": [
+        128,
+        94
+      ],
+      "B": [
+        74,
+        23
+      ],
+      "C": [
+        32,
+        12
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    }
+  },
+  "simulation_wall_s": 271.2116083000001
+}

docs/experiment_round51_k16pre.json

@@ -0,0 +1,110 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 16,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round51_k16pre_rows.json",
+    "out": "docs/experiment_round51_k16pre.json",
+    "trace": null,
+    "select": "pre",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "pre_h": 0,
+    "pre_near": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.56640625,
+    "collision_free_success": 0.56640625,
+    "time_ok_1p3": 0.359375,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.414156198501587,
+    "time_mean_s": 2.265103578567505,
+    "brakes_mean": 14.1953125,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.13815215229988098,
+    "err_fail_med": 184.40463256835938,
+    "err_success_med": 0.2881776690483093,
+    "dls_share": 0.64453125,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.2373737394809723,
+    "reverse_mean": 0.2401755303144455,
+    "reverse_moving_p50": 0.27848100662231445,
+    "moving_share_p50": 0.9851852059364319,
+    "brake_rate_mean": 0.07279647886753082,
+    "brake_rate_p50": 0.0,
+    "brake_rate_p90": 0.27435898780822754,
+    "stuck_share": 0.171875,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.596707820892334,
+    "by_cat": {
+      "A": [
+        128,
+        100
+      ],
+      "B": [
+        74,
+        26
+      ],
+      "C": [
+        32,
+        12
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    }
+  },
+  "simulation_wall_s": 486.8652108999959
+}

docs/experiment_round52_ctl_ex.json

@@ -0,0 +1,110 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round52_ctl_ex_rows.json",
+    "out": "docs/experiment_round52_ctl_ex.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "pre_h": 0,
+    "pre_near": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 0,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.5234375,
+    "collision_free_success": 0.5234375,
+    "time_ok_1p3": 0.265625,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.615795850753784,
+    "time_mean_s": 2.486567258834839,
+    "brakes_mean": 19.5546875,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.27920001745224,
+    "err_fail_med": 204.71360778808594,
+    "err_success_med": 0.21751396358013153,
+    "dls_share": 0.61328125,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.24741201102733612,
+    "reverse_mean": 0.25991496443748474,
+    "reverse_moving_p50": 0.32581019401550293,
+    "moving_share_p50": 0.9587156176567078,
+    "brake_rate_mean": 0.10028044879436493,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.3153846263885498,
+    "stuck_share": 0.234375,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5514403581619263,
+    "by_cat": {
+      "A": [
+        128,
+        96
+      ],
+      "B": [
+        74,
+        20
+      ],
+      "C": [
+        32,
+        11
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    }
+  },
+  "simulation_wall_s": 296.3142095999938
+}

docs/experiment_round53_ctl_ex_s1.json

@@ -0,0 +1,110 @@
+{
+  "config": {
+    "ckpt": "ml/runs/base_dag2_dag1/best.pt",
+    "samples": "ml/data/samples",
+    "tasks": null,
+    "split": "val",
+    "n": 256,
+    "K": 8,
+    "commit": 2,
+    "fb": 2.2,
+    "ki": 0.0,
+    "fb_scale": 0.0,
+    "fb_near": 0.2,
+    "fb_fallback": 0.0,
+    "kdecay": 0.9,
+    "kimax": 0.5,
+    "fb_every": 1,
+    "safety_try": 0,
+    "sel_w": 0.1,
+    "dls_safety": 1,
+    "fb_clear": 0.0,
+    "fb_cmin": 0.3,
+    "sub_refresh": 0,
+    "fb_block_gain": 0.5,
+    "dump": "ml/runs/round53_ctl_ex_s1_rows.json",
+    "out": "docs/experiment_round53_ctl_ex_s1.json",
+    "trace": null,
+    "select": "endpoint",
+    "replan": "empty",
+    "safety": 1,
+    "partial": 1,
+    "inc_cap": 1,
+    "gap_margin": 0.0,
+    "cert_batch": 1,
+    "steps": 1,
+    "z_scale": 1.0,
+    "len_w": 0.0,
+    "tie": 0,
+    "pre_h": 0,
+    "pre_near": 0.0,
+    "backoff": 0,
+    "max_sec": null,
+    "seed": 1,
+    "dls": "analytic",
+    "device": "cuda"
+  },
+  "batch_size": 256,
+  "batch_seed_rule": "same seed per batch",
+  "checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
+  "tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
+  "kin_fp": "2ebcbf17",
+  "act_scale": 1.0,
+  "metrics": {
+    "n": 256,
+    "n_ref": 256,
+    "success": 0.53125,
+    "collision_free_success": 0.53125,
+    "time_ok_1p3": 0.296875,
+    "collision_rate": 0.0,
+    "collision_steps_mean": 0.0,
+    "travel_ratio_median": 2.636941432952881,
+    "time_mean_s": 2.4600000381469727,
+    "brakes_mean": 19.265625,
+    "start_fail": 0.0,
+    "goal_ratio_2s": 0.2750342786312103,
+    "err_fail_med": 214.6193084716797,
+    "err_success_med": 0.22276456654071808,
+    "dls_share": 0.62890625,
+    "time_budget": "teacher×2(≥3 s)",
+    "speed_ratio_p50": 1.0000016689300537,
+    "speed_ratio_p95": 1.0000101327896118,
+    "speed_ratio_max": 1.0000101327896118,
+    "over_vmax_rate": 0.0,
+    "accel_ratio_p95": 1.0000016689300537,
+    "accel_ratio_max": 1.0000101327896118,
+    "reverse_p50": 0.25339367985725403,
+    "reverse_mean": 0.2575649917125702,
+    "reverse_moving_p50": 0.3117443919181824,
+    "moving_share_p50": 0.9636363387107849,
+    "brake_rate_mean": 0.09879808127880096,
+    "brake_rate_p50": 0.0051282052882015705,
+    "brake_rate_p90": 0.3384615480899811,
+    "stuck_share": 0.25390625,
+    "n_plans_mean": 195.0,
+    "success_reachable": 0.5596708059310913,
+    "by_cat": {
+      "A": [
+        128,
+        96
+      ],
+      "B": [
+        74,
+        21
+      ],
+      "C": [
+        32,
+        12
+      ],
+      "D": [
+        13,
+        0
+      ],
+      "Dr": [
+        9,
+        7
+      ]
+    }
+  },
+  "simulation_wall_s": 290.3290144000057
+}

docs/本机执行清单.md

@@ -3,51 +3,243 @@
 本文件是 Hark 对本机计算的唯一要求清单。按顺序执行，每一步把「回传」里的文件或输出原样交给用户，不要跳步，不要自行改代码。
 所有命令在仓库根目录（G:\Robot_v2\机器人手臂）的 Git Bash 里运行。完成的条目由 Hark 在下一次提交里删掉。
 
-当前状态（2026-10-19 02:00）：第二次预训练开环已正常（候选多样性 0.65、8 候选最好 2.9 mm），闭环仍 0% 到达 / 100% 碰撞；
-DAgger 阶段曾把显存吃到约 30 GB，原因是 kin.gaps 一次几千个点、cap=40 mm，已修复（kin.gaps 自动分块），尚未实测。
+## 已知结论（截至 2026-10-19 05:35 回传）
+
+- 现有模型 `ml/runs/base/last.pt` 闭环 0% 到达：1 段采样碰撞 99.6%；4 段采样 94%；4 段 + 仿真安全层碰撞 0.78%，到达仍 0%。consist / value / first 三种选法结果几乎一样，问题出在候选本身，不在挑哪条。
+- 三份代码审查逐行核对了滚动索引、前缀、提交、标签与执行的对齐，没有找到错位；但之前没有任何测试真正走过 `rollout()`。下面第 1 步的 H 段把「仿真错了」「前缀照抄」「子目标切换规则」「多峰」几种解释逐个分开。
+- 之前「前 1 s 速度比 2.31×」按 6 轴角度直接相加，被 J5/J6 抖动和 pos 任务冻结的 J6 放大。H 段改为分轴、净位移、TCP 三种口径。
 
 ## 第 0 步：同步代码
 
 ```bash
-git fetch gitea && git reset --hard gitea/main
-git log --oneline -1          # 应为「fix: 闭环失败定位与显存修复……」或更新
+git status --short            # 有未提交的改动先 commit（git add -A && git commit -m "本机改动"），不要丢
+git pull --rebase gitea main
+git log --oneline -1          # 应是 Hark 的「diag: H 段（老师 oracle / 前缀消融 / 子目标规则 / 分轴指标）…」
 ```
 
 - 训练控制台若开着，先点「停止」，确认没有 python 训练进程在跑。
-- 不要启动 DAgger / M5，等 Hark 看完第 1、2 步的结果再定。
+- 不要启动训练、DAgger、M5。
 
-## 第 1 步：gaps 显存峰值（不到 1 分钟）
-
-```bash
-PYTHONUTF8=1 python - <<'EOF'
-import torch
-from ml.common.kin_torch import Kin
-kin = Kin(device='cuda')
-for n, cap in [(512, 40.0), (4096, 4.0), (4096, 40.0)]:
-    q = kin.lo + (kin.hi - kin.lo) * torch.rand(n, 6, device='cuda')
-    torch.cuda.empty_cache(); torch.cuda.reset_peak_memory_stats()
-    with torch.no_grad(): kin.gaps(q, cap)
-    print(f'n={n:5d} cap={cap:4.0f}  peak {torch.cuda.max_memory_allocated() / 2**30:6.2f} GiB', flush=True)
-EOF
-```
-
-- 通过：三行峰值都 ≤ 2 GiB。
-- 任务管理器里专用显存接近 24 GB 或共享显存开始上涨：立刻 Ctrl+C，回传已打印的行，不要继续第 2 步。
-- 回传：终端输出的 3 行。
-
-## 第 2 步：闭环失败定位（约 2–4 分钟）
+## 第 1 步：闭环失败定位 H 段（约 35–50 分钟）
 
 ```bash
 PYTHONUTF8=1 python -m ml.eval.diag_cl --ckpt ml/runs/base/last.pt --n 256 --out ml/runs/diag_cl.json
 ```
 
-- 运行期间看一眼任务管理器，专用显存应远低于 24 GB；超过 22 GB 就 Ctrl+C 并回传终端输出。
-- 回传：`ml/runs/diag_cl.json`、`ml/runs/replay_teacher.json`、`ml/runs/base/log.jsonl`。
+- 专用显存超过 22 GB 就 Ctrl+C，回传终端最后 30 行和已写出的 `ml/runs/diag_cl.json`（每段跑完就写一次，中断也保留已完成的段）。
+- 现有 checkpoint 没有 EMA、没有 `act_scale`：脚本按原始权重、动作尺度 1.0 载入（输出里 `weights`=`model`、`act_scale_ckpt`=1.0），与之前各轮同口径。
+- 回传：`ml/runs/diag_cl.json`。
+
+### 输出键
+
+| 键 | 内容 |
+|---|---|
+| `A_*`、`B_act_diff_deg` | 起点特征 Python 与数据集逐组差；两份特征下的动作差（与之前相同） |
+| `C` | 旧部署口径（1 段采样、子目标旧规则 timeout）的失败形态。到达与碰撞改为只计每题自己的时限内，数字可能比上一轮略低 |
+| `H1_oracle` | 老师当策略走完整 `rollout()`。`oracle` 字段：`max_dev_deg_before_dls`（DLS 接管前与老师轨迹的最大关节偏差，p50 / p99 / max）、`reach`、`collision`、`finish_minus_teacher_s`（完成时刻 − 老师时长）、`dls_used` |
+| `H1_oracle_noise_0.05` / `_0.1` / `_0.2` | 老师动作每次规划加 N(0, σ²)（σ 以 STEP 为单位） |
+| `H2_open_loop` | r = 3 的 val 行：`s{1,4}_prefix_orig`（前缀原样，执行窗口 3..5 的误差，另含 `prefix_copy_err_deg_mean`）vs `s{1,4}_prefix_zeroed`（前缀置零） |
+| `H2_replan_empty_s1` / `_s4` | 队列空了才规划（每次 r = 0），其余同新部署默认 |
+| `H3_headsub_cross_s{1,4}` | 子目标头 + 越过顶点规则（新部署默认） |
+| `H3_headsub_timeout_s{1,4}` | 子目标头 + 旧规则（< 5° 或 1.5 s 超时，旧部署） |
+| `H3_teachersub_cross_s{1,4}` / `H3_teachersub_timeout_s{1,4}` | 老师真实途经点全表，两种切换规则 |
+| `H3_nosub_s{1,4}` | 不用子目标 |
+| `H4_open_loop` | 按前缀长度 r = 0..3 分组的执行窗口误差 \|Σ(act − Y)·STEP\|（度），1 段与 4 段采样 |
+
+每个闭环变体（H1–H3）都带同一组指标：`reach`、`collision`（cat ≠ D、时限内）、`goal_ratio_2s_p50`、`end_goal_err_mm_p50`、`first_hit_s_p50`、`t_done_over_teacher_p50`、`brakes_p50`、`joint_amp_ratio`（6 轴前 1 s 关节行程 策略/老师）、`joint_net_ratio`（6 轴前 1 s 净位移 策略/老师）、`j6_pos_amp_deg_p50`、`tcp_path_ratio_p50`、`tcp_net_ratio_p50`、`sign_flip_rate` 与 `sign_flip_rate_teacher`（6 轴前 2 s 相邻 3 步块变号比例）、`by_cat_mode`（各类别 n / reach / collision / goal_ratio_2s_p50）、`D`（D 类单列）。`_s1` = 1 段采样（部署），`_s4` = 4 段采样。
+
+### 判读
+
+| 现象 | 结论 |
+|---|---|
+| `H1_oracle` 的 reach < 0.95、collision > 0.01、或 `max_dev_deg_before_dls.p99` > 0.1° | 仿真 / 滚动本身有错，之前所有闭环 0% 都不可信。先修仿真，不看下面各项 |
+| `H1_oracle` 全部正常 | 仿真没问题，失败在策略输出。再对照 `H1_oracle_noise_*`：σ = 0.05 时 reach 就大幅下降，说明闭环对残差噪声极其敏感，单靠降噪不够 |
+| `H2_open_loop` 的 `prefix_zeroed` 执行窗口误差明显小于 `prefix_orig`，或 `H2_replan_empty` 明显好于同段数的 `H3_headsub_cross`（reach > 0，或 `goal_ratio_2s_p50` 明显更低、`sign_flip_rate` 明显更低） | 前缀照抄（copycat）是主因 |
+| `H3_teachersub_cross` 明显好于 `H3_teachersub_timeout`（或 `H3_headsub_cross` 好于 `H3_headsub_timeout`） | 子目标切换规则是主因之一 |
+| `H3_teachersub_*` 明显好于 `H3_headsub_*` | 子目标头给的途经点不可用 |
+| `H3_nosub` 不比 `H3_headsub_cross` 差 | 子目标通道可以整体去掉 |
+| `by_cat_mode` 里 `A/pos`（老师就是一条直线）reach 也是 0 | 失败与多峰无关 |
+| `joint_amp_ratio` 远大于 `joint_net_ratio`，且 `sign_flip_rate` 远高于 `sign_flip_rate_teacher` | 「动得多、走得少」是来回抖，不是速度律学快了 |
+
+重新打包和重训（`FRESH_FROM=prep`）要等 Hark 看完这一步的结果再定，这一步之后不要自行启动。
 
 ## 回传汇总
 
 | 步骤 | 回传内容 |
 |---|---|
-| 1 | 终端 3 行峰值 |
-| 2 | ml/runs/diag_cl.json、ml/runs/replay_teacher.json、ml/runs/base/log.jsonl |
+| 1 | ml/runs/diag_cl.json |
 | 任一步出错 | 终端最后 30 行输出 |
+
+---
+
+## 执行回传（2026-10-19 19:55，本地 AI）
+
+### 第 0 步：同步代码
+
+`ml/eval/diag_collision.py` 归档提交（本机改动 `7e063eb`）；根目录用户个人 html 为个人文件未入仓，原位保留。log 顶部 = `406f180`「diag: H 段 …」✓。无训练进程。
+
+### 第 1 步：H 段（完整 JSON 已更新到 ml/runs/diag_cl.json；weights=model、act_scale_ckpt=1.0，同口径；H_tasks 256、missing_teacher 0）
+
+**判读表逐行核对：**
+
+1. **`H1_oracle`：仿真完全正常。** reach **1.0**（243/243）、collision **0.0**、`max_dev_deg_before_dls` p50 6.9e-5° / p99 4.6e-4°（阈值 0.1°）、`finish−teacher` p50 −0.02 s、`dls_used` 94.9%。→ **滚动/仿真没有错，之前所有闭环 0% 都可信，失败在策略输出本身。**
+2. **`H1_oracle_noise`：对残差噪声极其敏感。** σ=0.05 时 reach 1.0→**0.218**（碰撞 9.1%）；σ=0.1→0.058；σ=0.2→0.012。→ 单靠降采样噪声不够（判读表第二行命中）。
+3. **`H2_open_loop`：前缀照抄排除。** prefix_orig vs zeroed：s1 2.315° vs **2.375**（置零不更好）；s4 0.251° vs **0.705**（置零明显更差）。`H4` 按 r=0..3 分组误差持平（2.31–2.33°）。`H2_replan_empty`：s1 0%/79.4%/0.836、s4 0%/84.8%/0.603——不比同段数 `H3_headsub_cross` 好。
+4. **`H3`：子目标三个变体逐位相同（新异常）。** `headsub_cross` / `headsub_timeout` / `nosub` 在 s1、s4 下从 reach 到 end_err 到 flip 数组**完全一致**（s1: 0% / 96.3% / 0.7504 / 414.2091；s4: 0% / 79.8% / 0.5869 / 428.0032）。`teachersub_*`（真值途经点）仅微动（s1 95.9%/0.726、s4 79.4%/0.578）。→ **子目标头通道在这条闭环里零响应**——比「可以整体去掉」更强；可能是 need 概率全程 <0.5 从不触发，或子目标输入对旧模型输出无影响，建议核一下（D0 曾测过 need 标签占比 28.1%）。
+5. **`by_cat_mode`：A/pos（老师直线）reach 也是 0**（s1 碰撞 94%）→ **失败与多峰无关**。
+6. **分轴指标：s1 的"动得多"含抖、s4 后基本消抖。** J5/J6 幅度/净位移：s1 3.08/2.57、2.55/2.16；s4 1.83/1.84、0.69/0.69，`sign_flip_rate` 从 0.026–0.062 降到 0.0006–0.0013（≈老师 0.0005–0.004 水平）。→ 4 段采样把来回抖消掉了，但净位移仍 J5 1.8×，且到达仍 0%。
+
+**C 段**（新时限口径）：0% 到达 / 96.5% 碰撞 / 2 s 误差比 0.755。
+
+### 本轮无出错，无终端报错段。
+
+---
+
+## 本地自主排查（2026-10-19 20:40，用户授权「继续排查」）
+
+### 1. 闭环噪声容忍阈值（老师动作 + N(0,σ²)，σ 以 STEP 为单位，n=256，其余同 H1_oracle）
+
+| σ | 0 | 0.005 | 0.01 | 0.02 | 0.03 | 0.04 | 0.05 |
+|---|---|---|---|---|---|---|---|
+| reach | 1.0 | 1.0 | 0.971 | 0.716 | 0.457 | 0.309 | 0.218 |
+| collision | 0 | 0.8% | 2.5% | 4.9% | 6.2% | 7.4% | 9.1% |
+| ratio2s | 0.047 | 0.049 | 0.049 | 0.054 | 0.059 | 0.063 | — |
+
+→ **闭环能容忍的每步噪声上限 ≈ 0.01–0.02×STEP**（约 0.02–0.04°/步）。策略 s1 残差 ~0.77°/步、s4 ~0.05–0.08°/步，都在崩溃区右侧。
+
+### 2. 8 段采样闭环（steps=8，同 H3_headsub_cross 口径）
+
+reach **0%** / collision **79.0%** / ratio2s 0.574 / end 408 mm；`sign_flip_rate` 0.0004–0.0015（≈老师 0.0005–0.004，抖已完全消掉）、`amp ≈ net`（1.00/1.19/0.97/1.19/1.36/0.76，无抖动成分）。
+
+→ 对比 s4（reach 0 / 79.8% / 0.587）：**s4→s8 几乎无改善**。按上表的白噪声口径，s8 的残差幅度（16 步 0.56° ≈ 0.035°/步，σ_eff≈0.0175）应给出 reach≈0.75——实测 0。**策略误差不是白噪声，是有方向的系统偏差**：幅度不大但同向、闭环里逐步累积，重规划抵消不掉。降采样噪声（1→4→8 段）已到天花板。
+
+### 3. 子目标通道：need 头恒不触发（新发现，修正「子目标排除」的理由）
+
+- 冷启动（子目标特征抹零，与闭环同口径）need 概率：mean 0.0048、p50 0.0030、p90 0.0079、**max 0.0819**、>0.5 占比 **0**；43 道 need=True 题召回 **0/43**。（D0 的 0.71875 准确率恰等于"全预测 False"基线，当时已有此暗示。）
+- 因此闭环里 `sub_flag` 恒 0、子目标从不激活——这才是 H3 三变体逐位相同的直接原因。
+- 但模型对子目标特征**高度敏感**（同一噪声下真值子目标 vs 置零：|Δact| mean 0.177、max 3.09）——训练时 80% 样本 sf=1，而闭环 100% 是 sf=0，属训练/闭环分布错配。
+- 不过 H3_teachersub（强制真值 need + 真值途经点）仍 0% / 95.9%——**子目标不是闭环全灭的主因**，但 need 头失效 + 闭环恒走少数分支值得修。
+
+### 4. 汇总判断
+
+- 仿真正常（H1_oracle 已证）；采样噪声已压到老师抖动水平仍 0% 到达 → **剩余障碍是策略的系统性动作偏差 + 闭环对误差的极低容忍度（误差复利）**。
+- 继续调采样段数、子目标规则、候选选择都无解；正确方向是**在闭环状态上训练（DAgger）**，让老师纠正分布偏移，或提高闭环鲁棒性。
+- 未自行启动 DAgger / 重训，等指示。
+
+### 5. 补充：系统偏差的方向结构（4096 行开环 + 256 题闭环，4 段采样）
+
+**开环动作误差（deg，有符号）**：
+
+- 每轴均值 [0.010, −0.012, −0.001, −0.029, 0.034, 0.018]，|mean|/std ≤ 0.17 → **无固定轴向偏置**（不是某轴恒偏）。
+- 每轴标准差 0.16–0.21°；每步平均 |误差| 0.016–0.018°，**16 步内不随步数增长**。
+- **lag-1 自相关 0.931–0.967（极高）** → 误差是**慢变/连贯**的，不是白噪声。这解释了为什么白噪声 σ 扫描会低估策略误差的危害：闭环重规划抵消不掉持续同向的偏差。
+- 误差与老师动作方向：cos mean 0.046 / p50 0.072（不超调也不欠调）；投影比 e·Y/|Y|² mean −0.038 / p50 0.006（无系统性缩放）。
+
+**闭环误差增长（s4，n=256，done=0）**，goal_err / 起点误差：
+
+| t (s) | 0.25 | 0.50 | 0.75 | 1.00 | 1.50 | 2.00 |
+|---|---|---|---|---|---|---|
+| p50 | 0.959 | 0.880 | 0.811 | 0.741 | 0.645 | 0.591 |
+| p90 | 1.024 | 1.072 | 1.119 | 1.162 | 1.222 | **1.318** |
+
+→ p50 缓慢靠近（2 s 缩 41%），**p90 持续发散**——≥10% 的题误差随时间增长（误差复利进碰撞轨道），其余也远不够快。
+
+**补充结论**：策略每步误差很小（~0.017°）但**时间相关性极强（0.93–0.97）**，闭环对"慢变误差"的敏感度远超等幅白噪声；失败是「慢变误差 × 闭环误差复利」。推理侧降噪（段数/选择器）已无空间，训练侧需要降低慢变误差（DAgger / 闭环状态训练）。
+
+### 6. 「2 s 误差比 >1 的发散题」能否提前筛掉（per-task 分组，s4，n=256；明细 ml/runs/task_divergence.json）
+
+| 组（2 s 误差比） | 题数 | 碰撞 | 到达 | cat/mode | 起点误差 p50 | 老师时长 p50 | need 率 | 老师到达 |
+|---|---|---|---|---|---|---|---|---|
+| 发散 r2>1.0 | 48 | 81% | 0% | A/pos 14、A/pose 14、C/pose 5、C/pos 5、B/pos 3、B/pose 3、D/pos 2、Dr/pos 2 | 374 mm | 1.80 s | 21% | 96% |
+| 平 0.7–1.0 | 48 | 83% | 0% | 混合 | 539 mm | 2.12 s | 27% | 90% |
+| 收敛 ≤0.7 | 160 | 79% | 0% | 混合 | 642 mm | 2.18 s | 31% | 96% |
+
+- **三组碰撞率相同（79–83%）、到达全 0**——不存在"一小撮可提前去掉的坏题"；发散题散布在所有类别，老师都能到达。
+- r2 与起点误差秩相关 **−0.384**（起点误差小 → 比值易大），是**分母效应**：短任务（老师 1.8 s、e0 374 mm）比值波动大，不代表题目特殊；r2 与老师时长/行程几乎无关（−0.09 / −0.03）。
+- 结论：误差放大是**全体题目上的普遍行为**，筛题无效（只让指标好看），必须训练侧修正。
+
+### 7. DAgger 小样本试跑（用户授权；未跑全量、未微调）
+
+**链路实测**（ckpt `ml/runs/base/last.pt`，旧口径 act_scale=1.0）：
+
+| 步骤 | 规模 | 实测 |
+|---|---|---|
+| `ml.dagger.collect`（16 题，parts=2） | 16 题 | **12.7 s**（含加载），29 个起点（全撞，拒 1 个 gap） |
+| `ml.dagger.collect --n 64`（显存采样） | 64 题 | 成功，112 个起点（拒 10 gap），**显存峰值 7.98 GB**（24 GB 卡，无 OOM） |
+| `node ml/gen/gen_dataset.cjs --tasks tasks_01.jsonl` | 15 个起点 | **7.6 s**，shard 含 qPrev / v0 / traj / cert |
+| `ACT_SCALE=1.0 node ml/prep/build_dataset.cjs …` | 13 题 | **2.3 s**，453 样本（175 增广），meta `act_scale:1, data_ver:2` |
+
+**关键口径提醒**：`build_dataset.cjs` 现在默认 **ACT_SCALE=1.25**（DATA_VER 2），而现有主数据 `ml/data/samples` 无 meta → 训练按 **1.0** 读。两者不一致时训练直接报错（README 第 120 行）。全量 DAgger 前二选一：
+1. 打包 DAgger 时显式 `ACT_SCALE=1.0`（与旧主数据、旧 ckpt 兼容）；
+2. 先按新流程 `FRESH_FROM=prep` 重打主数据为 1.25、重训后再跑 DAgger。
+本次试跑全部用 1.0；产物在 `ml/data/dag_try*/`（gitignored）。
+
+**全量耗时外推（估算，以首次实测为准）**：collect 8000 题约 0.8–1 小时（单进程）；老师重规划 1.6 万起点 / NPROC≈12 约 10 分钟；打包约 3 分钟；微调 15000 步未试。
+
+**结论**：DAgger 收集链路可用、显存安全（峰值 8 GB），可以全量启动——但先定 act_scale 口径。
+
+### 8. 「加大数据能否力挽狂澜」——闭环状态到训练分布的距离（s4，n=256，9408 个闭环状态）
+
+方法：闭环轨迹每 10 步取一个 q（9408 点），对训练集 191 万老师状态求最近邻（6 维关节角 L∞，±360 回绕）；对照 = 每题起点（老师分布内）。
+
+| 状态 | p10 | p50 | p90 | p99 |
+|---|---|---|---|---|
+| 闭环轨迹状态 | 12.10° | **21.46°** | 32.50° | 44.56° |
+| 每题起点（老师） | 8.71° | 13.21° | 17.91° | 21.53° |
+
+- 闭环状态离最近老师状态中位 **21.5°**（起点 13.2° 的 1.6×；p90 32.5°、p99 44.6°）。
+- 训练集本身在关节空间稀疏：连老师起点最近邻都有 13.2°——数据是"轨迹线"而非密集体。
+- 按密度缩放（6 维 d∝N^(−1/6)）：数据翻 4 倍最近邻仅降 ~21%（21.5°→~17°），而策略误差尺度是 0.02°；差距三个数量级，且同类数据不改变误差的时间相关结构（0.94）。
+- **结论：扩大同类老师数据救不了闭环**；需要「不同的数据」——在策略实际到达的状态上标注（DAgger），把 21.5° 这些分布外点变成训练样本。
+
+### 9. 方案 A 实测：合成扰动状态样本微调（两轮配方）
+
+做法：从训练任务老师轨迹的 15–85% 处取状态，加 N(0,σ²) 关节扰动（σ 混合 0.5/1/2/4°；实测起点偏移 p50 1.5°），老师重规划（**2000/2000 到达**、cert 1951/2000），打包 36k 样本（act_scale=1.0，为主数据的 1.9%），从 `base/last.pt` 微调 3000 步，2000 步时以 EMA 权重做 512 题闭环评估。
+
+| 配方 | success | collision | goal_ratio_2s | ol_fk_mm | ol_best_mm | ol_div |
+|---|---|---|---|---|---|---|
+| base/last.pt（基线） | 0 | 0.996 | 0.755 | 5.98 | 2.92 | 0.655 |
+| +扰动样本 lr 3e-4 | 0 | **0.484** | 0.927 | **15.27** | 10.32 | 0.966 |
+| +扰动样本 lr 1e-4 | 0 | 0.604 | 0.893 | **16.42** | 11.45 | 0.980 |
+
+- 碰撞率首次大幅下降（99.6%→48–60%），但 **success 仍 0**；开环精度退化 2.5–2.7 倍（6→15/16 mm）、多样性升高——小步微调在两种分布间折中，把主分布带偏了。
+- **结论**：1.5° 量级的合成扰动 + 3000 步微调不能解决闭环。若继续这条线，需从**预训练开始**按课程混入扰动样本（并监控开环不退化），成本高；或转向结构改动（方案 B：绝对目标参数化/参考轨迹跟踪）或真实 DAgger。
+- 产物：`ml/data/samples_perturb/`、`ml/runs/base_perturb/`、`ml/runs/base_perturb2/`（均在 gitignored 目录）。
+
+### 10. 方案 B 判别实验：模型预测的"累计目标"是否指向参考轨迹
+
+方法：2000 个扰动起点（原轨迹状态 + σ 0.5–4° 扰动，L∞ 实测 p50 **3.19°**），模型预测 16 步增量并累积为绝对目标 `Z_model = q_off + cumsum(â·STEP)`，对比原老师轨迹未来 16 步绝对位置 `Z_ref`；`ρ = |Z_model − Z_ref| / |q_off − q_ref|`（ρ≈1 = 目标随状态平移、无纠偏；ρ≈0 = 目标稳定指向轨迹）。
+
+| 模型 | 状态偏移 p50 | 目标误差 p50 | ρ p50 |
+|---|---|---|---|
+| base/last.pt | 3.19° | 4.86° | **1.54** |
+| base_perturb（扰动微调 lr 3e-4） | 3.19° | 5.85° | 1.90 |
+| base_perturb2（lr 1e-4） | 3.19° | 6.25° | 1.97 |
+| 对照：未扰动行开环 | — | **0.46°** | — |
+
+- 未扰动时 16 步累计目标误差仅 0.46°；状态偏移 3.19° 后目标偏离 4.86°（ρ=1.54）——**当前模型是"从当前状态外推增量"，不是"指向参考位置"**。扰动微调让 ρ 更差（1.9–2.0），与开环退化一致。
+- **结论：方案 B 不能靠改执行层实现，必须改训练标签**——把标签从"16 步增量"改为"16 步绝对目标关节角（参考轨迹位置）"，让模型学"目标在哪"而不是"从这怎么走"。执行层每步朝目标走，偏移状态下即天然纠偏。
+- 实施成本：改 `build_dataset.cjs` 标签口径（Y = 绝对目标），同步 FK/平滑/价值等基于增量的辅助损失、rollout 与 `ik-policy.js` 的增量假设，然后从头训练（约 1 小时）。属结构性改动，未擅自开始。
+
+### 11. 方案 B 初版实测（分支 `feat/abs-target`，保存点 tag `pre-b-abs-target`）
+
+**改动**（全部在分支，main 未动）：
+- `build_dataset.cjs`：`ABS_TARGET=1` → Y = 16 步绝对目标关节角（/180 归一化），prefix 仍存增量（度），meta 记录 `abs_target`、签名加 `a1`。
+- `train.py`：读 `abs_target`；每单位动作=180°；FK 损失用绝对位置；平滑项换算到增量量纲；开环指标用第 16 步绝对位置；ckpt 记录 `abs_target`。
+- `closed_loop.py`：模型输出绝对目标 → 队列存目标序列、执行 `q=clamp(target)`、前缀由目标差分、安全层中点与 DLS 增量转绝对目标。
+- `ckpt.py` / `data.py`：abs_target 透传与一致性检查。
+- **过程 bug（已修）**：前缀噪声增广把 eps 直接加到绝对目标标签上（≈18° 污染，一半样本）；改为累计偏移后第 r 步扣回。
+
+**数据**：84 shard 重打 → 201 万样本、abs_target=1；标签连续（步间差 p50 0.30°、p99 1.8°、无 ±180 跳变）。
+
+**训练结果**（从零 9950 步；p_neq 预热 5000 步）：
+
+| step | ol_fk_mm | ol_best_mm | 闭环 success | 碰撞 |
+|---|---|---|---|---|
+| 4000 | 74.6 | 33.4 | 0 | 100% |
+| 6000 | 140.4 | 70.3 | 0 | 100% |
+| 8000 | 139.0 | 68.5 | 0 | 100% |
+
+- 一步采样动作误差 **31°**（4/8 段采样 2.3°/2.8°）——**绝对目标的一步流匹配映射没学会**；闭环 1 段 0%/100%、4 段 0%/100%（ratio2s 1.045）。
+- 过预热后 5000 步无下降趋势。可能原因：一步样本仅 `p_edge=10%`、`ada_p=1` 逆误差加权压制难样本、FK/平滑辅助项在绝对位置空间主导。
+- **结论：此配方不成立**。B 若继续，需改训练配方（p_edge↑、w_fk↓、ada_p=0）或换输出头（确定性回归）；当前证据不支持继续投入。分支代码保留可复现。

ml/README.md

@@ -24,7 +24,7 @@
 | 3 老师门禁 | `bash ml/pipeline.sh gate1` | 25–30 分钟（单核） | `ml/data/consistency.json` |
 | 4 生成数据 | `bash ml/pipeline.sh gen` | 不共识约 1 小时；共识约 4–5 小时 | `ml/data/shard_*.jsonl`、`stats.json` |
 | 5 打包样本 | `bash ml/pipeline.sh prep` | 15–30 分钟 | `ml/data/samples/`、`ml/data/r2.json` |
-| 6 预训练 | `bash ml/pipeline.sh train` | 测速 2000 步约 6 分钟，正式约 3 小时 | `ml/runs/base/best.pt` |
+| 6 预训练 | `bash ml/pipeline.sh train` | 测速 2000 步约 6 分钟，正式 24000 步约 1.5–2.5 小时（未实测：fp32 MF 分支后的吞吐以测速为准） | `ml/runs/base/best.pt` |
 | 6′ DAgger 纠偏 | `bash ml/pipeline.sh dagger` | 未实测，估 3–4 小时：2 轮 ×（采起点 5 分钟 + 老师重规划约 30 分钟 + 打包 + 微调 15000 步） | `ml/runs/base_dagger.ckpt`、`ml/runs/base_dag2/best.pt` |
 | 7 M5 门禁 | `bash ml/pipeline.sh m5` | 估算 10–20 分钟 | `ml/runs/m5_gate_base.json` |
 | 8 APG 微调 | `bash ml/pipeline.sh apg` | 未实测：展开长度 60–150 次推理随批变化，以监控页前 50 步换算的 ETA 为准 | `ml/runs/apg/best.pt` |
@@ -115,7 +115,9 @@ bash ml/pipeline.sh gen                     # 默认 5 万条、进程数 = CPU
 bash ml/pipeline.sh prep
 ```
 
-每条轨迹按 50 Hz 时间律重放，每 3 步取一个样本（特征与部署同一份 JS 代码计算），30% 样本加 σ = 1° 噪声增广（加噪后间隙 < 3 mm 的不用）。
+每条轨迹按 50 Hz 时间律重放，每 3 步取一个样本（特征与部署同一份 JS 代码计算）。40% 样本对 q 加 σ = 3° 噪声增广，标签改为 12 步内线性收回老师轨迹（加噪后间隙 < 3 mm 或收回途中有碰撞的不用）。
+- **动作标签**：Y = 每步增量 / (act_scale · STEP)，夹在 ±1，`act_scale` = 1.25（`ACT_SCALE`）。老师巡航时主导轴每步正好 1 个 STEP，标签最大 0.8，执行端的 ±1 只防越界；被 ±1 截掉的收回量在下一步补上。运行时增量 = clamp(act, ±1) · act_scale · STEP，`act_scale` 写在 checkpoint 和 ONNX meta 里，仿真、APG、浏览器运行时都按它反归一化；没有这个字段的旧模型按 1.0 处理。
+- 每个 shard 打包产物带 `<shard>.meta.json`（数据版本 `data_ver`=2、`act_scale`、增广参数、签名 `sig`）。`ACT_SCALE`、`AUG_P`、`AUG_SIGMA`、`AUG_BACK`、`MIN_GAP_MM` 任一变化，签名就变，`run_prep.sh` 只重打签名不符的 shard；没有 meta 的旧打包一律重打。训练读 meta 取 `act_scale`，同一次训练的各目录（主数据 + DAgger）必须一致，否则报错。
 认证最小间隙 < 3 mm 的老师轨迹不进训练集（`MIN_GAP_MM=3`，与 `stats.json` 的 `gapDrop` 同一口径）：模型只学离障碍 ≥ 3 mm 的走法，运行时投影护栏仍保持 5 mm。
 检查 `ml/data/samples/index.json`：`train_samples` 约 170 万、`val_samples` 约 9 万（5% 任务划为 val）。
 `MIN_GAP_MM` 必须与生成阶段相同（`stats.json` 的 `minGapFilterMm`），不一致时这一步直接报错（返回码 8）。
@@ -128,15 +130,22 @@ bash ml/pipeline.sh prep
 bash ml/pipeline.sh train                           # MeanFlow 头 → ml/runs/bench_base/、ml/runs/base/
 ```
 
-开训前先跑老师轨迹回放自检（`ml/runs/replay_teacher.json`，约 1 分钟）：老师轨迹放进 PyTorch 仿真口径，到达 ≥ 95%、碰撞 ≤ 1% 才开训，否则 rc 5 停——口径对不上时闭环指标不可信。
-闭环指标长期 0% 而开环（ol_fk_mm / ol_best_mm / ol_div）正常时，跑 `PYTHONUTF8=1 python -m ml.eval.diag_cl --ckpt ml/runs/base/last.pt --n 256`（约 2–4 分钟，可与训练并行）→ `ml/runs/diag_cl.json`：A 起点特征 Python 与 JS 逐组差、B 两份特征下动作差、C 首次碰撞时刻 / 组对 / 最小目标误差 / 开安全层到达率。
+开训前先跑老师轨迹回放自检（`ml/runs/replay_teacher.json`，约 1–3 分钟），两部分都过才开训：
+- 运动学口径：老师 50 Hz 轨迹直接放进 PyTorch 运动学 / 碰撞 / 到达判据，到达 ≥ 95%、碰撞 ≤ 1%；
+- 滚动口径：老师当策略（`TeacherOracle`）走完整的闭环仿真 `rollout()`（队列、前缀、提交 3 步、标签 clamp、DLS、时限都与策略同一条路径），时限内到达 ≥ 95%、碰撞 ≤ 1%、DLS 接管前与老师轨迹的最大关节偏差 p99 ≤ 0.1°。
+流程读 json 的 `pass` 字段，并核对 `fp`（当前 `kin.json`）与 `data_fp`（数据 shard 记录的碰撞模型指纹）；任一不符就重跑一次回放，仍不过 rc 5 停。
 
-先跑 2000 步测吞吐（`ml/runs/bench_base/log.jsonl` 的 `samples_per_s`，`eta_h` 是剩余小时数），再正式训练到 60000 步或闭环到达率连续 5 次不升为止。
+闭环指标长期 0% 而开环正常时，跑 `PYTHONUTF8=1 python -m ml.eval.diag_cl --ckpt ml/runs/base/last.pt --n 256`（约 35–50 分钟）→ `ml/runs/diag_cl.json`：A 起点特征 Python 与 JS 逐组差、B 两份特征下动作差、C 旧部署口径的失败形态、H 老师 oracle / 前缀消融 / 子目标规则 / 分轴与分类别指标（各键含义见 `ml/eval/diag_cl.py` 文件头与 `docs/本机执行清单.md`）。
 
-- **通过线（R10）：`samples_per_s` ≥ 5000，即 60000 步 ≤ 3.5 小时**。参考：3090、batch 1024 实测 6.2 k 样本/s ≈ 2.8 小时。MeanFlow 的 JVP 双前向使训练受算力限制，加大 batch 不会提高样本/s；低于通过线时把 `configs/base.json` 的 `fk_caps` 从 96 降到 48，仍不够再把 `steps` 降到 40000
-- 训练中每 2000 步打印一行 `EVAL {...}`：val 任务闭环（部署节奏、不开安全层）`success`、`collision_rate`、`time_ok_1p3`、`start_fail`。时限 = 老师时长 × 2（至少 3 s）；另有开环三项（val 样本、老师状态上）：`ol_fk_mm` 候选 0 的 16 步末端误差、`ol_best_mm` 8 个候选里最好的一条、`ol_div` 候选多样性（候选间标准差 / 动作幅度，塌成单峰时接近 0）。best.pt 与早停按「到达率 − 0.001 × ol_fk_mm」
-- 损失（2026-10-18 首次正式预训练后改）：FK / 平滑只加在 MF 分支 r = t 且 t ≤ 0.3 的配对去噪估计 x̂ = z − t·u 上（原来加在随机噪声的一步采样上，等于回归条件均值，多峰塌成单峰）；时间嵌入最高角频率 50（原 1000）；MF 分支与 JVP 走 fp32（TF32 矩阵乘）；前 5000 步 p_neq = 0。fp32 会让吞吐低于原 6.2 k/s，R10 以新实测为准
-- 数据增广：40% 样本加 σ = 3° 噪声、12 步收回（原 30% / 1° / 8 步），收回途中须无碰
+先跑 2000 步测吞吐（`ml/runs/bench_base/log.jsonl` 的 `samples_per_s`，`eta_h` 是剩余小时数），再正式训练 24000 步（`configs/base.json` 的 `steps`，余弦学习率按这个预算走完）。
+
+- **通过线（R10）：`samples_per_s` ≥ 5000**（24000 步 × batch 1024 ≈ 1.4 小时）。MeanFlow 的 JVP 双前向使训练受算力限制，加大 batch 不会提高样本/s；低于通过线时把 `configs/base.json` 的 `fk_caps` 从 96 降到 48
+- **EMA**：训练同时维护权重的指数滑动平均（`ema` 0.999），checkpoint 里存 `model`（原始权重）与 `ema`。评估、DAgger、APG 起点、M5、挖错、验收、导出一律用 EMA 权重（没有 `ema` 的旧 checkpoint 用 `model`）
+- 训练中每 2000 步用 EMA 权重打印一行 `EVAL {...}`：val 任务闭环（部署节奏、不开安全层）`success`、`collision_rate`、`goal_ratio_2s`（2 s 时目标误差 / 起点误差的中位数）、`time_ok_1p3`、`start_fail`，时限 = 老师时长 × 2（至少 3 s），碰撞与到达只计时限内；开环三项（val 样本、老师状态上）：`ol_fk_mm` 候选 0 的 16 步末端误差、`ol_best_mm` 8 个候选里最好的一条、`ol_div` 候选多样性。
+- **选模**：`score` = success + 0.05·(1 − collision_rate) + 0.05·(1 − min(1, goal_ratio_2s)) − 1e-4·min(ol_best_mm, 500)，到达率全 0 时闭环指标仍能排序。p_neq 预热期内和预热后第一次评估之前不存 best.pt（一步映射还没训过）；早停只在 max(预热 + 2 次评估, 总步数一半) 之后计数，连续 5 次不升即停
+- 损失：FK / 平滑只加在 MF 分支 r = t 且 t ≤ 0.3 的配对去噪估计 x̂ = z − t·u 上；时间嵌入最高角频率 50；MF 分支与 JVP 走 fp32（TF32 矩阵乘）；前 5000 步 p_neq = 0，之后 p_neq = 0.5，另有 10% 样本固定在部署用的 (r = 0, t = 1)（`p_edge`）
+- 子目标头只用把 sub 特征置零的输入训练（sub 存在时子目标标签与输入是同一个量，原样抄就能答对）；运行时子目标头只在冷启动、sub 为空时使用，口径一致。价值头用完整输入
+- 条件输入增广（训练时现做，`aug_pre_drop` 0.15 / `aug_pre_sigma` 0.1 / `aug_dq_sigma` 0.3）：前缀随机清空（变成 r = 0 样本，标签不变）；前缀加噪声，标签前 r 步同步加、第 r 步扣回；速度特征 dq 加噪声。目的是让模型不能靠照抄上一块和上一帧速度出动作
 
 ## 6′. DAgger 纠偏
 
@@ -144,7 +153,8 @@ bash ml/pipeline.sh train                           # MeanFlow 头 → ml/runs/b
 bash ml/pipeline.sh dagger
 ```
 
-用预训练 best.pt 在 train 任务上闭环跑（关安全层），每题取「首次碰撞前 0.2 s」和一个随机时刻的状态做起点（须无碰、间隙 > `MIN_GAP_MM`、离目标 > 5 mm），老师从这些起点重新规划（`gen_dataset.cjs --tasks`，与主数据同认证、同时间律），打包后与主数据一起微调。默认 2 轮（`DAG_ROUNDS`），每轮 8000 题（`DAG_N`）、微调 15000 步（`DAG_STEPS`，lr 3e-4）。老师只用于训练标注，部署仍无老师。M5 门禁读 `ml/runs/<头>_dagger.ckpt` 指向的模型；APG / 挖错的 BC 项也用 DAgger 样本。可断点续跑（采集、老师 shard、打包、微调各有完成标记）。
+起点模型：预训练 best.pt 的闭环 success > 0 时用它，否则用 last.pt（`python -m ml.common.ckpt pick ml/runs/<头>`），都取 EMA 权重；之后每轮按同一规则取上一轮的结果。微调时只有起点已走过自己的 p_neq 预热才不再预热。
+用起点模型在 train 任务上闭环跑（关安全层），每题取「首次碰撞前 0.2 s」和一个随机时刻的状态做起点（须无碰、间隙 > `MIN_GAP_MM`、离目标 > 5 mm），同时记下该时刻的上一帧。老师从这些起点重新规划（`gen_dataset.cjs --tasks`，与主数据同认证），时间律从当前速度在首段方向上的分量起步（`v0`），t = 0 样本的速度特征用真实上一帧；打包后与主数据一起微调。微调的 val 只取主数据，DAgger 样本全部算 train。默认 2 轮（`DAG_ROUNDS`），每轮 8000 题（`DAG_N`）、微调 15000 步（`DAG_STEPS`，lr 3e-4）。老师只用于训练标注，部署仍无老师。M5 门禁读 `ml/runs/<头>_dagger.ckpt` 指向的模型；APG / 挖错的 BC 项也用 DAgger 样本。可断点续跑（采集、老师 shard、打包、微调各有完成标记）。
 - 显存预算**浮动**：整卡留 1 GB 给系统（`VRAM_RESERVE_GIB`，默认 1），训练与 APG 进程启动时按"整卡 − nvidia-smi 已用 − 余量"和"CUDA 空闲 − 余量"两者取小自设本进程上限（`ml/common/vram.py`）。占显存的应用多，上限就自动变小，不用改配置；想按定值对拍给 `VRAM_FRAC`（旧口径 0.78）。日志 `vram` 字段记峰值；预训练 `batch=1024` 实测峰值 2.76 GiB、APG `batch=48` 13.08 GiB。显存不够把 `batch` 改成 512。中断后再跑 `bash ml/pipeline.sh train`（或 `all`）会自动从 `ml/runs/<头>/last.pt` 续训（每 10 分钟存一次），测速那 2000 步做过就跳过。训练发散时不会用发散的权重覆盖 `last.pt` / `best.pt`。
 
 ## 7. M5 门禁：选定动作头
@@ -191,7 +201,7 @@ bash ml/pipeline.sh mine            # 默认 6 轮；ROUNDS=3 APG_STEPS=1000 MIN
 ## 10. 导出
 
 ```bash
-bash ml/pipeline.sh export          # fp32 ONNX + meta.json → webapp/models/（约 20 MB）
+bash ml/pipeline.sh export          # fp32 ONNX + meta.json（含 act_scale）→ webapp/models/（约 20 MB），导出 EMA 权重
 ```
 
 默认 fp32：部署走 wasm 与 Node CPU，fp16 在 CPU 后端上不保证更快。只有第 13 步用 `&ep=webgpu` 实测 `__POLICY.latency()` 明显受益时，才另导出 fp16 做对比：
@@ -234,6 +244,7 @@ python ml/serve.py                  # 带跨源隔离头的本地服务，wasm
 - 控制台 `__POLICY.latency()` 查看推理、前瞻、安全层、整次规划的 P50/P99 与节拍；`__POLICY.stats()` 查看刹停、换向、精修次数
 - `&autotest=30` 自动跑与 bench 同一批贴柱题，结果在 `__POLICY.autotest`
 - 运行时投影护栏 5 mm：候选动作离障碍 < 5 mm 时沿间隙梯度外推（训练数据的过滤线是 3 mm，护栏比数据更保守）
+- 动作尺度读模型 meta 的 `act_scale`；子目标按「越过顶点」切换（与训练标签同口径：沿拐角平分面越过当前途经点即换下一个，没有超时），与仿真 `closed_loop.sub_advance` 同一规则
 - 工具偏移 ≠ 0 时自动交回原规划器
 - 离线环境：`bash ml/prep/vendor_ort.sh` 把 onnxruntime-web 放到 `webapp/vendor/ort/`，地址加 `&ortLocal=1`（WebGPU 同样可用）
 
@@ -314,7 +325,8 @@ ml/
   sim/closed_loop.py       部署节奏闭环仿真（验证 / 挖错 / 门禁）
   apg/apg_finetune.py      APG 可微微调
   mining/                  tasks.py（无老师出题）、mine_round.py、gate.py（含 R7）、loop.sh
-  eval/                    dump_parity.cjs + check_parity.py（JS↔PyTorch 对拍）、check_r2.py（R2）、r4_probe.py（R4 无头浏览器探针）、acceptance.py（验收汇总）
+  eval/                    dump_parity.cjs + check_parity.py（JS↔PyTorch 对拍）、check_r2.py（R2）、r4_probe.py（R4 无头浏览器探针）、acceptance.py（验收汇总）、
+                           replay_teacher.py（开训前回放自检，含老师 oracle 走 rollout）、diag_cl.py（闭环失败定位 A/B/C/H）
   toy/toy_bimodal.py       2D 双模态对照实验（动作头选型依据）
   export_onnx.py  serve.py
   configs/                 base.json（MeanFlow）、wta.json（回退）、small.json（R4 不过时的 d=128）、apg.json

ml/apg/apg_finetune.py

@@ -6,7 +6,10 @@
 2026-10-18：展开长度不再定长 60 次推理（3.6 s，会截掉老师时长 > 3.6 s 的绕远题），改为覆盖本批 exec_s p90×1.2，上限 plans_max；
   "到达"仍是每 seg_plans 段末的距离项（无到达即停），属有意设计：段末项给出持续的接近梯度。
   margin_train 3 mm：与训练数据过滤线一致（认证最小间隙 < 3 mm 的老师轨迹不进训练集），碰撞项不与 BC 标签对抗；
-  运行时投影护栏独立保持 5 mm。端点与中点两个铰链取均值，与只查端点同量纲，w_col 无需重标。"""
+  运行时投影护栏独立保持 5 mm。端点与中点两个铰链取均值，与只查端点同量纲，w_col 无需重标。
+2026-10-19：起点取 EMA 权重（ckpt.load_policy）；动作反归一化乘 act_scale（与数据同一尺度，不一致报错），产物存 act_scale；
+  子目标推进改用 closed_loop.sub_advance（默认「越过顶点」，与标签同口径）；子目标头 BC 项用 sub 特征置零的输入（同 train.py）；
+  val 只取主数据目录（第一个 --samples）。"""
 import argparse, json, math, os, random, sys, time
 import torch
 
@@ -15,7 +18,7 @@ from ml.common.features import features, STEP, H, wrap180
 from ml.data import Samples, load_tasks
 from ml.model.policy import build, wta_loss
 from ml.model.meanflow import meanflow_loss
-from ml.sim.closed_loop import evaluate, tasks_to_tensors, goal_err
+from ml.sim.closed_loop import evaluate, tasks_to_tensors, goal_err, sub_advance
 from ml.common.ckpt import Saver, atomic_save, rng_state, set_rng_state
 from ml.common import vram
 
@@ -41,18 +44,23 @@ def main():
     dev = a.device; torch.manual_seed(a.seed); random.seed(a.seed)
     vram.limit()
     os.makedirs(a.out, exist_ok=True)
-    ck = torch.load(a.init, map_location=dev)
-    mcfg = ck['cfg']
-    model = build(mcfg).to(dev); model.load_state_dict(ck['model'])
+    from ml.common.ckpt import load_policy
+    model, ck = load_policy(a.init, dev); mcfg = ck['cfg']
     kin = Kin(device=dev)
-    data = Samples(a.samples, device=dev, split='train')
+    data = Samples(a.samples, device=dev, split='train', val_dirs=a.samples[:1])
+    act_scale = float(ck.get('act_scale', 1.0))
+    if abs(act_scale - data.act_scale) > 1e-9:
+        raise SystemExit(f'起点 act_scale={act_scale} 与样本 act_scale={data.act_scale} 不一致（BC 项标签尺度不同），换同一版数据训练的模型')
+    model.act_scale = act_scale
     all_tasks = load_tasks(a.samples)
-    train_tasks = [t for t in all_tasks if t['split'] == 'train'] + load_pool(a.pool)
-    val_tasks = [t for t in all_tasks if t['split'] == 'val'][:C['eval_tasks']]
+    main_tid = {t['tid'] for t in load_tasks(a.samples[:1]) if t['split'] == 'val'}
+    train_tasks = [t for t in all_tasks if t['tid'] not in main_tid] + load_pool(a.pool)
+    val_tasks = [t for t in load_tasks(a.samples[:1]) if t['split'] == 'val'][:C['eval_tasks']]
     w = [C['cat_weights'].get(t.get('cat', 'A'), 1.0) * (2.0 if t.get('hard') else 1.0) for t in train_tasks]
     opt = torch.optim.AdamW(model.parameters(), lr=C['lr'], betas=(0.9, 0.95), weight_decay=0.0)
     amp = dev.startswith('cuda') and C.get('amp') == 'bf16'
     step = torch.tensor(STEP, device=dev)
+    astep = step * act_scale                                   # 动作 → 度
     it0 = 0
     if a.resume and os.path.exists(a.resume):
         rk = torch.load(a.resume, map_location=dev); model.load_state_dict(rk['model'])
@@ -68,11 +76,11 @@ def main():
         base, _ = evaluate(model.eval(), kin, val_tasks, dev)
         print('EVAL0', json.dumps(base), flush=True); log.write(json.dumps({'eval': base, 'step': 0}) + '\n')
         best = base['success']
-        atomic_save({'model': model.state_dict(), 'cfg': mcfg, 'step': 0, 'fp': kin.fp, 'eval': base}, os.path.join(a.out, 'best.pt'))
+        atomic_save({'model': model.state_dict(), 'cfg': mcfg, 'step': 0, 'fp': kin.fp, 'eval': base, 'act_scale': act_scale}, os.path.join(a.out, 'best.pt'))
     M = C['margin_train_mm']; t0 = time.time()
     saver = Saver(C.get('ckpt_min', 10))             # 2026-10-18：每 10 分钟 + 收到停止请求时存 last.pt（不再只在每 500 次验证时存）
     def save_last(n, ev=None):
-        atomic_save({'model': model.state_dict(), 'opt': opt.state_dict(), 'cfg': mcfg, 'step': n, 'fp': kin.fp, 'eval': ev, 'rng': rng_state()}, os.path.join(a.out, 'last.pt')); saver.mark()
+        atomic_save({'model': model.state_dict(), 'opt': opt.state_dict(), 'cfg': mcfg, 'step': n, 'fp': kin.fp, 'eval': ev, 'rng': rng_state(), 'act_scale': act_scale}, os.path.join(a.out, 'last.pt')); saver.mark()
     for it in range(it0, C['steps']):
         model.train()
         batch = random.choices(train_tasks, weights=w, k=C['batch'])
@@ -81,7 +89,7 @@ def main():
         z = torch.randn(B, H, 6, device=dev)
         q, qp = q0.clone(), q0.clone()
         queue = torch.zeros(B, 0, 6, device=dev)
-        subs = torch.zeros(B, 2, 6, device=dev); nsub = torch.zeros(B, dtype=torch.long, device=dev); sub_t = torch.zeros(B, device=dev)
+        subs = torch.zeros(B, 2, 6, device=dev); nsub = torch.zeros(B, dtype=torch.long, device=dev); sub_t = torch.zeros(B, device=dev); sub_prev = q0.clone()
         opt.zero_grad(set_to_none=True)
         stats = {'col': 0.0, 'tgt': 0.0, 'smooth': 0.0}
         # 展开长度随本批任务时长：覆盖本批 exec_s 的 p90 × 1.2（无老师参考题按 6 s），夹在 [plans, plans_max]
@@ -95,11 +103,7 @@ def main():
             pre = torch.zeros(B, 3, 6, device=dev); pre[:, :r] = queue[:, :r]
             qe = q + queue.sum(1)
             with torch.no_grad():
-                if (nsub > 0).any():
-                    cur = subs[:, 0]; dd = torch.where(kin.cont, wrap180(cur - qe), cur - qe).abs().max(-1).values
-                    adv = (nsub > 0) & ((dd < 5) | (tnow - sub_t > 1.5))
-                    subs = torch.where(adv[:, None, None], torch.cat([subs[:, 1:], subs[:, 1:]], 1), subs)
-                    nsub = torch.where(adv, nsub - 1, nsub); sub_t = torch.where(adv, torch.full_like(sub_t, tnow), sub_t)
+                subs, nsub, sub_prev, sub_t = sub_advance(kin, subs, nsub, sub_prev, sub_t, qe.detach(), tnow)
             feat = features(kin, q, qp, T, R, pre, torch.full((B,), r, device=dev), subs[:, 0], (nsub > 0).float())
             with torch.autocast('cuda', dtype=torch.bfloat16, enabled=amp):
                 act, val, sub = model.sample_k(feat, 1, z)
@@ -108,8 +112,8 @@ def main():
                 with torch.no_grad():
                     need = sub[:, 12].float() > 0.5
                     subs = q.detach()[:, None] + sub[:, :12].float().view(B, 2, 6) * 180
-                    nsub = torch.where(need, torch.full_like(nsub, 2), torch.zeros_like(nsub))
-            inc = act.clamp(-1, 1) * step                              # [B,16,6]
+                    nsub = torch.where(need, torch.full_like(nsub, 2), torch.zeros_like(nsub)); sub_prev = q.detach().clone()
+            inc = act.clamp(-1, 1) * astep                             # [B,16,6]
             commit = inc[:, r:r + 3]
             if r == 0:   # 冷启动：先提交一块，下一轮才开始执行
                 queue = commit; continue
@@ -118,7 +122,7 @@ def main():
                 d = queue[:, k]
                 qp = q; q = kin.clamp_box(q + d)
                 # 碰撞铰链：端点 + 中点（与仿真/运行时的离散检查一致）
-                G = kin.gaps(torch.cat([q, 0.5 * (q + qp)], 0), M)
+                G = kin.gaps(torch.cat([q, 0.5 * (q + qp)], 0), M, pairs=True)
                 hinge = torch.relu(M - G['gap'])
                 col = torch.zeros(2 * B, device=dev).index_add(0, G['b'], hinge).view(2, B).mean(0)   # 端点与中点取均值：与只查端点时同量纲，w_col 无需重标
                 sm = ((d / step) ** 2).sum(-1) + (((d - prev_d) / step) ** 2).sum(-1)
@@ -135,13 +139,18 @@ def main():
                 stats['tgt'] += lt.mean().item()
                 (seg_loss / nseg).backward()
                 seg_loss = torch.zeros((), device=dev)
-                q, qp, queue, prev_d = q.detach(), qp.detach(), queue.detach(), prev_d.detach()
+                # 只 detach 状态并重置队列：段边界下一 plan 走 r=0 重新提交（新图有梯度）；
+                # 沿用旧 queue 会跨过已 backward 的图（二次 backward 报错），也避免整段无梯度
+                q, qp, prev_d = q.detach(), qp.detach(), prev_d.detach()
+                queue = torch.zeros(B, 0, 6, device=dev)
         # BC 锚定
         X, Y, V, SG = data.batch(C['batch'] * 4)
         with torch.autocast('cuda', dtype=torch.bfloat16, enabled=amp):
             mem = model.encode(X)
             l_bc = wta_loss(model, mem, Y)[0] if model.head == 'wta' else meanflow_loss(model, mem, Y)[0]
-            v, s = model.heads(mem)
+            v, _ = model.heads(mem)
+            Xs = X.clone(); Xs[:, 68:75] = 0
+            _, s = model.heads(model.encode(Xs))
         l_bc = l_bc + 0.1 * (v.float() - V).abs().mean() + 0.1 * (s.float() - SG).abs().mean()
         (C['w_bc'] * l_bc).backward()
         gn = torch.nn.utils.clip_grad_norm_(model.parameters(), C['grad_clip'])
@@ -160,7 +169,7 @@ def main():
             res, _ = evaluate(model.eval(), kin, val_tasks, dev)
             res['it'] = it + 1; res['t'] = time.time()
             print('EVAL', json.dumps(res), flush=True); log.write(json.dumps({'eval': res}) + '\n'); log.flush()
-            sd = {'model': model.state_dict(), 'cfg': mcfg, 'step': it + 1, 'fp': kin.fp, 'eval': res}
+            sd = {'model': model.state_dict(), 'cfg': mcfg, 'step': it + 1, 'fp': kin.fp, 'eval': res, 'act_scale': act_scale}
             save_last(it + 1, res)                                                      # last 带优化器状态供续跑；best 不带（导出/门禁用）
             if res['success'] > best + 1e-4:
                 best = res['success']; atomic_save(sd, os.path.join(a.out, 'best.pt'))

ml/common/ckpt.py

@@ -60,3 +60,33 @@ def set_rng_state(d):
         try: torch.cuda.set_rng_state_all([t.cpu() for t in d['cuda']])
         except Exception: pass
     return True
+
+
+def load_policy(path, device, prefer_ema=True):
+    """载入策略（2026-10-19，审查 review_model_train P0-2）：优先取 EMA 权重 ck['ema']（train.py 起存），没有就用 ck['model']（旧 checkpoint、APG 产物）。
+    model.act_scale = ck['act_scale']（动作标签归一化倍数，见 ml/prep/build_dataset.cjs ACT_SCALE）；旧 checkpoint 缺省 1.0 = 原口径。
+    闭环仿真 / DAgger / APG 起点 / 挖错 / 门禁 / 验收 / 导出都经这里载入，保证同一份权重、同一个动作尺度。返回 (model.eval(), ck)。"""
+    from ml.model.policy import build
+    ck = torch.load(path, map_location=device)
+    m = build(ck['cfg']).to(device)
+    sd = ck.get('ema') if prefer_ema and ck.get('ema') is not None else ck['model']
+    m.load_state_dict(sd)
+    m.act_scale = float(ck.get('act_scale', 1.0))
+    return m.eval(), ck
+
+
+def pick_start(run_dir):
+    """DAgger 起点（2026-10-19，审查 P0-2）：best.pt 的闭环 success > 0 才用 best.pt；否则用 last.pt
+    （success 全 0 时 best.pt 只由开环误差决定，可能落在 p_neq 预热期，一步映射没训过）。载入时取 EMA。"""
+    best, last = os.path.join(run_dir, 'best.pt'), os.path.join(run_dir, 'last.pt')
+    if os.path.exists(best):
+        ev = torch.load(best, map_location='cpu').get('eval') or {}
+        if (ev.get('success') or 0) > 0 or not os.path.exists(last): return best
+    return last if os.path.exists(last) else best
+
+
+if __name__ == '__main__':
+    # python -m ml.common.ckpt pick <运行目录>   → 打印 DAgger 起点 checkpoint 路径
+    import sys
+    if len(sys.argv) == 3 and sys.argv[1] == 'pick': print(pick_start(sys.argv[2]))
+    else: raise SystemExit('usage: python -m ml.common.ckpt pick <run_dir>')

ml/common/features.py

@@ -52,8 +52,10 @@ def features(kin, q, q_prev, T, R, prefix, prefix_len, sub, sub_flag, gm=None, x
     out.append(sd)
     out.append(sub_flag[:, None].to(dt))
     if gm is None:
-        gm = kin.gaps(q, GAP_CAP, xf)['gm']
-    out.append(torch.stack([gm[:, a * 7 + b].clamp(max=GAP_CAP) / GAP_CAP for a, b in GP], -1))
+        out.append(kin.gm_groups(q, GP, GAP_CAP, xf=xf) / GAP_CAP)   # 2026-10-20：只算特征用到的 8 组；xf 复用不重复算 FK
+        gm = None
+    if gm is not None:
+        out.append(torch.stack([gm[:, a * 7 + b].clamp(max=GAP_CAP) / GAP_CAP for a, b in GP], -1))
     f = torch.cat(out, -1)
     assert f.shape[1] == F_DIM, f.shape
     return f

ml/common/kin_torch.py

@@ -38,7 +38,8 @@ def seg_seg_dist(p1, q1, p2, q2, eps=1e-12):
     s = torch.where(t_lo, (-c / a.clamp_min(eps)).clamp(0, 1), s)
     s = torch.where(t_hi, ((b - c) / a.clamp_min(eps)).clamp(0, 1), s)
     t = t.clamp(0, 1)
-    # 退化（线段长度≈0）
+    # 第二条退化成点时，仍须在第一条上投影，而不是固定取其起点。
+    s = torch.where(e <= eps, (-c / a.clamp_min(eps)).clamp(0, 1), s)
     s = torch.where(a <= eps, torch.zeros_like(s), s)
     t = torch.where(e <= eps, torch.zeros_like(t), t)
     c1 = p1 + d1 * s.unsqueeze(-1)
@@ -85,8 +86,36 @@ class Kin:
         gi, gj = self.cap_g[self.pi], self.cap_g[self.pj]
         self.pair_gp = torch.minimum(gi, gj) * 7 + torch.maximum(gi, gj)
         self.lever = T(K['lever'])
+        table = [[0.0] * 7 for _ in range(6)]
+        for group in range(1, 7):
+            endpoints = [c[key] for c in caps if c['g'] == group for key in ('a', 'b')]
+            if not endpoints:
+                continue
+            pivot, axis = K['P'][group - 1], K['D'][group - 1]
+            unit = [v / math.hypot(*axis) for v in axis]
+            relative = [[e[j] - pivot[j] for j in range(3)] for e in endpoints]
+            radius = max(math.hypot(*v) for v in relative)
+            radial = max(math.hypot(v[1] * unit[2] - v[2] * unit[1],
+                                    v[2] * unit[0] - v[0] * unit[2],
+                                    v[0] * unit[1] - v[1] * unit[0]) for v in relative)
+            for joint in range(group):
+                chain = sum(math.dist(K['P'][j], K['P'][j + 1]) for j in range(joint, group - 1))
+                table[joint][group] = (chain + (radial if joint == group - 1 else radius)) * D2R
+        self.cert_groups = torch.tensor([a * 7 + b for a in range(7) for b in range(a + 2, 7)], device=device)
+        # 三角不等式给出全姿态界，不用随机姿态最大值冒充上界。
+        self.cert_lever = T([[table[k][b] if a <= k < b else 0.0 for k in range(6)]
+                             for a in range(7) for b in range(a + 2, 7)])
 
     # ---------------- 运动学 ----------------
+    def _checked_xf(self, q, xf):
+        """xf 是按行对应 q 的位姿数组，行数不匹配会静默广播出错（宽相位筛到别的行 ⇒ device assert 或错间隙）。"""
+        if xf is None:
+            return self.cum_xf(q)
+        Rs, Ts = xf
+        if Rs.shape[0] != q.shape[0] or Ts.shape[0] != q.shape[0]:
+            raise ValueError('xf 的行数 %d 与 q 的行数 %d 不符' % (Rs.shape[0], q.shape[0]))
+        return xf
+
     def cum_xf(self, q):
         """q: [B,6] 度 → Rs [B,7,3,3], Ts [B,7,3]"""
         B = q.shape[0]
@@ -110,6 +139,19 @@ class Kin:
         Rs, _ = xf if xf is not None else self.cum_xf(q)
         return Rs[:, 6]
 
+    def jac(self, q, xf=None):
+        """几何雅可比（对度）：返回 (Jv [B,3,6] mm/deg, Jw [B,3,6] 单位轴/deg)"""
+        Rs, Ts = xf if xf is not None else self.cum_xf(q)
+        p = self.fk(q, (Rs, Ts))
+        Jv, Jw = [], []
+        for k in range(6):
+            d = torch.tensor(self.D[k], dtype=q.dtype, device=q.device)
+            a = Rs[:, k] @ d
+            o = torch.einsum('bij,j->bi', Rs[:, k], self.P[k].to(q.dtype)) + Ts[:, k]
+            Jv.append(torch.linalg.cross(a, p - o, dim=-1) * D2R)
+            Jw.append(a * D2R)
+        return torch.stack(Jv, -1), torch.stack(Jw, -1)
+
     def caps_world(self, q, xf=None):
         Rs, Ts = xf if xf is not None else self.cum_xf(q)
         Rg, Tg = Rs[:, self.cap_g], Ts[:, self.cap_g]               # [B,C,3,3], [B,C,3]
@@ -118,24 +160,53 @@ class Kin:
         return a, b
 
     # ---------------- 间隙 ----------------
-    # 2026-10-19（用户实测 dagger/诊断吃满 24 GB 显存）：粗筛后幸存胶囊对数随 点数 × cap 暴涨，
-    # 一次几千个点、cap=40 mm 时中间张量可达十几 GB（训练里 eval 峰值 15.9 GiB 也是它）。
-    # 现在先按簇级粗筛估出每个点的幸存胶囊对数，按 PAIR_BUDGET 把 batch 切块逐块算，峰值与调用方的 batch 大小无关。
+    # 2026-10-19（用户实测 dagger/诊断吃满 24 GB 显存）：本模型 1936 条胶囊 / 491 组 / 67256 粗对 / 104.7 万细对。
+    # 显存有两段峰值，都随点数线性涨：
+    #   粗筛段 cc[:,cp_a]-cc[:,cp_b] 的广播中间量 ≈ 36 B × 点数 × 粗对数（4096 点实测 9.3 GiB，与 cap 无关）；
+    #   精算段 ≈ 每对几百字节 × 幸存细对数（4096 点 cap 40 幸存 2.4 亿对；旧版不分块实测 36.9 GiB）。
+    # 所以先按 COARSE_ENTRIES 把点切块（粗筛每块 ≤ ~1 GiB），块内再按幸存细对数 PAIR_BUDGET 切（精算每块 ≤ ~1 GiB）。
+    # 两段不同时存在，峰值与调用方 batch 大小无关。结果与不分块逐位一致（梯度经 cat 保留）。
+    COARSE_ENTRIES = 30_000_000
     PAIR_BUDGET = 2_000_000
 
-    def gaps(self, q, cap=40.0, xf=None, max_pairs=None):
+    @staticmethod
+    def _cat_gaps(parts, starts):
+        out = {'gm': torch.cat([p.pop('gm') for p in parts], 0)}
+        if 'b' in parts[0]:
+            # 非 in-place：b 会被 autograd 的索引反向图引用（APG 的碰撞铰链梯度路径），+= 会让 backward 报 version 错
+            for p, s in zip(parts, starts): p['b'] = 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, 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)
+        xf = self._checked_xf(q, xf)
+        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, 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]), pairs) for s in st], st)
+
+    def _coarse(self, q, cap, xf):
+        Rs, Ts = xf
+        cc = torch.einsum('bkij,kj->bki', Rs[:, self.cl_g], self.cl_c.to(q.dtype)) + Ts[:, self.cl_g]
+        d2 = (cc[:, self.cp_a] - cc[:, self.cp_b]).pow(2).sum(-1)          # [B,M]
+        del cc
+        return d2 <= (self.cp_lim + cap).pow(2)
+
+    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]
         with torch.no_grad():
-            cc = torch.einsum('bkij,kj->bki', Rs[:, self.cl_g], self.cl_c.to(q.dtype)) + Ts[:, self.cl_g]
-            d2 = (cc[:, self.cp_a] - cc[:, self.cp_b]).pow(2).sum(-1)
-            per = ((d2 <= (self.cp_lim + cap).pow(2)).long() * self.cp_cnt).sum(1)       # [B] 每点幸存胶囊对数
-            del cc, d2
+            per = (self._coarse(q, cap, xf).long() * self.cp_cnt).sum(1)    # [B] 每点幸存细对数
             if B <= 1 or int(per.sum()) <= self.PAIR_BUDGET:
                 bounds = [0, B]
             else:
@@ -144,22 +215,18 @@ class Kin:
                     if c - base > self.PAIR_BUDGET and k > bounds[-1]:
                         bounds.append(k); base = cs[k - 1]
                 bounds.append(B)
+            del per
         if len(bounds) == 2:
-            return self._gaps_one(q, cap, xf)
-        parts = [self._gaps_one(q[s:e], cap, (Rs[s:e], Ts[s:e])) for s, e in zip(bounds[:-1], bounds[1:])]
-        return {'gm': torch.cat([p['gm'] for p in parts], 0),
-                'b': torch.cat([p['b'] + s for p, s in zip(parts, bounds[:-1])]),
-                '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])}
+            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():
-            cc = torch.einsum('bkij,kj->bki', Rs[:, self.cl_g], self.cl_c.to(q.dtype)) + Ts[:, self.cl_g]
-            d2 = (cc[:, self.cp_a] - cc[:, self.cp_b]).pow(2).sum(-1)          # [B,M]
-            keep = d2 <= (self.cp_lim + cap).pow(2)
+            keep = self._coarse(q, cap, xf)
             bb, mm = keep.nonzero(as_tuple=True)
+            del keep
             cnt = self.cp_cnt[mm]
             rep_b = torch.repeat_interleave(bb, cnt)
             off = torch.repeat_interleave(self.cp_start[mm] - torch.cumsum(cnt, 0) + cnt, cnt) + torch.arange(int(cnt.sum()), device=q.device)
@@ -174,8 +241,130 @@ 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 gm_groups(self, q, gp, cap=40.0, xf=None):
+        """只算指定组对 gp=[(a,b),...] 的最小间隙 [B,len(gp)]（特征用，避免 49 组全量）。
+        数值与 gaps(q,cap)['gm'][:, a*7+b] 一致；组对在碰撞粗对里不存在时返回 cap。
+        xf 可传入调用者已算好的 cum_xf(q)，避免同一姿态重复算 FK（特征路径每周期省一次）。"""
+        key = tuple(gp)
+        if getattr(self, '_gp_key', None) != key:
+            m = torch.full((49,), -1, dtype=torch.long, device=q.device)
+            for i, (a_, b_) in enumerate(gp): m[min(a_, b_) * 7 + max(a_, b_)] = i
+            ga, gb = self.cl_g[self.cp_a], self.cl_g[self.cp_b]
+            cg = torch.minimum(ga, gb) * 7 + torch.maximum(ga, gb)   # 簇对 → 7 连杆组对
+            sel = torch.isin(cg, torch.tensor([min(a, b) * 7 + max(a, b) for a, b in gp], device=q.device))
+            self._gp = (m, sel, sel.nonzero(as_tuple=True)[0]); self._gp_key = key
+        self._gp49, sel, idx = self._gp
+        xf = self._checked_xf(q, xf); Rs, Ts = xf; B = q.shape[0]
+        with torch.no_grad():
+            cc = torch.einsum('bkij,kj->bki', Rs[:, self.cl_g], self.cl_c.to(q.dtype)) + Ts[:, self.cl_g]
+            d2 = (cc[:, self.cp_a[idx]] - cc[:, self.cp_b[idx]]).pow(2).sum(-1)
+            keep = d2 <= (self.cp_lim[idx] + cap).pow(2)
+            bb, mm = keep.nonzero(as_tuple=True)
+            del keep, d2, cc
+            cnt = self.cp_cnt[idx][mm]
+            rep_b = torch.repeat_interleave(bb, cnt)
+            off = torch.repeat_interleave(self.cp_start[idx][mm] - torch.cumsum(cnt, 0) + cnt, cnt) + torch.arange(int(cnt.sum()), device=q.device)
+            pidx = off
+        a, b = self.caps_world(q, xf)
+        with torch.no_grad():
+            pi, pj = self.pi[pidx], self.pj[pidx]
+            mid = 0.5 * (a[rep_b, pi] + b[rep_b, pi] - a[rep_b, pj] - b[rep_b, pj])
+            k2 = mid.pow(2).sum(-1) <= (self.pair_bl[pidx] + cap).pow(2)
+            rep_b, pidx, pi, pj = rep_b[k2], pidx[k2], pi[k2], pj[k2]
+        d = seg_seg_dist(a[rep_b, pi], b[rep_b, pi], a[rep_b, pj], b[rep_b, pj])
+        gap = d - self.pair_need[pidx].to(q.dtype)
+        gm = torch.full((B * len(gp),), float(cap), dtype=q.dtype, device=q.device)
+        gm = gm.scatter_reduce(0, rep_b * len(gp) + self._gp49[self.pair_gp[pidx]], gap.clamp(max=cap), reduce='amin', include_self=True)
+        return gm.view(B, len(gp))
+
+    @torch.no_grad()
+    def certify_segments(self, a, b, max_queries=32, max_depth=12, eps=1e-4, cap=20.0):
+        """逐层二分认证：每根段独立查询预算，未知（超预算/超深度/非有限/越界）一律拒绝。
+
+        判据对每条组对：子段两端间隙之和 > 该子段位移界 + 2eps（间隙取 min(gap,cap)，钳制只会更保守）。
+        cap 必须 ≥ 该层子段界 + 2eps 才不会挡住本可通过的判定，故不做自适应收紧——同批其他段会改变
+        钳制程度、让结果随批次而变（实测也无耗时收益，见探索日志轮次 31）。
+        """
+        if self.dtype != torch.float64 or a.dtype != torch.float64 or b.dtype != torch.float64:
+            raise ValueError('segment certification requires float64')
+        if a.shape != b.shape or a.ndim != 2 or a.shape[1] != 6:
+            raise ValueError('a and b must be [count,6] with identical shape')
+        # 预算给成张量时 max_depth + 1 会在 uint8 这类窄类型上回绕成 0，循环整个跳过，把碰撞段当安全返回。
+        if type(max_queries) is not int or type(max_depth) is not int:
+            raise ValueError('certification budgets must be plain ints')
+        if max_queries < 2 or max_depth < 0 or eps <= 0 or cap <= eps:
+            raise ValueError('invalid certification budget or tolerance')
+        count = a.shape[0]
+        status = torch.zeros(count, dtype=torch.long, device=a.device)
+        inside = (((a >= self.lo) & (a <= self.hi) & (b >= self.lo) & (b <= self.hi)) | self.cont).all(-1)
+        status[~inside] = 3
+        root = torch.arange(count, device=a.device)
+        gaps = self.gaps(torch.cat([a, b]), cap)['gm'][:, self.cert_groups]
+        ga, gb = gaps[:count], gaps[count:]
+        queries = torch.full((count,), 2, dtype=torch.long, device=a.device)
+        bound = (b - a).abs() @ self.cert_lever.T
+        for depth in range(max_depth + 1):
+            finite = torch.isfinite(ga).all(-1) & torch.isfinite(gb).all(-1)
+            collision = (ga <= 0).any(-1) | (gb <= 0).any(-1)
+            status[root[~finite]] = 2
+            status[root[collision]] = 1
+            positive = (ga > eps).all(-1) & (gb > eps).all(-1)
+            proven = positive & (ga + gb > bound + 2 * eps).all(-1)
+            pending = (status[root] == 0) & ~proven
+            if not bool(pending.any()):
+                break
+            a, b, ga, gb, root, bound = a[pending], b[pending], ga[pending], gb[pending], root[pending], bound[pending]
+            if depth == max_depth:
+                status[root] = 2
+                break
+            requested = torch.bincount(root, minlength=count)
+            status[(status == 0) & (queries + requested > max_queries)] = 2
+            keep = status[root] == 0
+            a, b, ga, gb, root, bound = (t[keep] for t in (a, b, ga, gb, root, bound))
+            if not root.numel():
+                break
+            mid = a + 0.5 * (b - a)
+            gm = self.gaps(mid, cap)['gm'][:, self.cert_groups]
+            queries += torch.bincount(root, minlength=count)
+            a, b = torch.cat([a, mid]), torch.cat([mid, b])
+            ga, gb = torch.cat([ga, gm]), torch.cat([gm, gb])
+            root = torch.cat([root, root])
+            bound = torch.cat([bound, bound]) / 2
+        return {'safe': status == 0, 'collision': status == 1, 'unresolved': status == 2,
+                'outside_box': status == 3, 'queries': queries}
+
+    @torch.no_grad()
+    def certify_uniform(self, a, b, n=64, eps=1e-4, cap=None):
+        """固定层数的均匀细分认证：n 个子段只需一次 gaps 调用（n+1 个点），没有逐层依赖的往返。
+
+        判据与 certify_segments 完全相同；界对 Δq 线性，所以子段界 = 整段界 / n。未通过即 unproven
+        （拒绝），不存在“未知当安全”。cap=None 时取“最小不挡住可通过判定的宽相位”= 最大子段界 + 2eps：
+        小于它钳制会把本可证明的子段压成不可证（偏保守），大于它只是多算对（偏慢），不改正确性。
+        """
+        if self.dtype != torch.float64 or a.dtype != torch.float64 or b.dtype != torch.float64:
+            raise ValueError('segment certification requires float64')
+        if type(n) is not int or n < 1 or eps <= 0 or (cap is not None and cap <= eps):
+            raise ValueError('invalid subdivision count or tolerance')
+        if a.shape != b.shape or a.ndim != 2 or a.shape[1] != 6:
+            raise ValueError('a and b must be [count,6] with identical shape')
+        count = a.shape[0]
+        inside = (((a >= self.lo) & (a <= self.hi) & (b >= self.lo) & (b <= self.hi)) | self.cont).all(-1)
+        bound = (b - a).abs() @ self.cert_lever.T / n                       # [count,15] 子段界
+        required = float(bound.max()) + 2 * eps if count else 2 * eps
+        qc = required if cap is None else cap
+        alpha = torch.arange(n + 1, dtype=self.dtype, device=a.device) / n
+        points = (a[:, None] + (b - a)[:, None] * alpha[None, :, None]).reshape(-1, 6)
+        gm = self.gaps(points, qc)['gm'][:, self.cert_groups].view(count, n + 1, -1)
+        adjacent = gm[:, :-1] + gm[:, 1:] > bound[:, None, :] + 2 * eps
+        collision = (gm <= 0).any((1, 2))
+        safe = inside & (gm > eps).all((1, 2)) & adjacent.all((1, 2)) & ~collision
+        return {'safe': safe, 'collision': collision, 'unproven': (~safe & ~collision & inside),
+                'outside_box': ~inside, 'points': n + 1, 'queries': n + 1,
+                'cap_required_mm': required, 'cap_used_mm': qc}
+
     def min_gap(self, q, cap=40.0):
         return self.gaps(q, cap)['gm'].min(-1).values
 

ml/configs/base.json

@@ -1,10 +1,11 @@
 {
   "model": {"d": 192, "heads": 8, "ff": 768, "enc": 4, "dec": 4},
   "train": {
-    "batch": 1024, "lr": 1e-3, "warmup": 1000, "steps": 60000, "wd": 0.01, "betas": [0.9, 0.95],
+    "batch": 1024, "lr": 1e-3, "warmup": 1000, "steps": 24000, "wd": 0.01, "betas": [0.9, 0.95],
     "amp": "bf16", "grad_clip": 1.0,
     "w_mf": 1.0, "w_fk": 0.5, "w_smooth": 0.05, "w_value": 0.1, "w_sub": 0.1,
-    "p_neq": 0.25, "p_neq_warm": 5000, "fk_tmax": 0.3, "fk_caps": 96,
+    "p_neq": 0.5, "p_edge": 0.1, "p_neq_warm": 5000, "fk_tmax": 0.3, "fk_caps": 96,
+    "ema": 0.999, "aug_pre_drop": 0.15, "aug_pre_sigma": 0.1, "aug_dq_sigma": 0.3,
     "ckpt_min": 10, "eval_every": 2000, "eval_tasks": 512, "log_every": 50, "patience_evals": 5
   }
 }

ml/configs/base_dag2.json

@@ -0,0 +1,11 @@
+{
+  "model": {"d": 192, "heads": 8, "ff": 768, "enc": 4, "dec": 4},
+  "train": {
+    "batch": 1024, "lr": 1e-3, "warmup": 1000, "steps": 24000, "wd": 0.01, "betas": [0.9, 0.95],
+    "amp": "bf16", "grad_clip": 1.0,
+    "w_mf": 1.0, "w_fk": 0.5, "w_smooth": 0.05, "w_value": 0.1, "w_sub": 0.1,
+    "p_neq": 0.5, "p_edge": 0.1, "p_neq_warm": 5000, "fk_tmax": 0.3, "fk_caps": 96,
+    "ema": 0.999, "aug_pre_drop": 0.15, "aug_pre_sigma": 0.1, "aug_dq_sigma": 0.3,
+    "ckpt_min": 10, "eval_every": 2000, "eval_tasks": 512, "log_every": 50, "patience_evals": 5
+  }
+}

ml/configs/base_dag3.json

@@ -0,0 +1,11 @@
+{
+  "model": {"d": 192, "heads": 8, "ff": 768, "enc": 4, "dec": 4},
+  "train": {
+    "batch": 1024, "lr": 1e-3, "warmup": 1000, "steps": 24000, "wd": 0.01, "betas": [0.9, 0.95],
+    "amp": "bf16", "grad_clip": 1.0,
+    "w_mf": 1.0, "w_fk": 0.5, "w_smooth": 0.05, "w_value": 0.1, "w_sub": 0.1,
+    "p_neq": 0.5, "p_edge": 0.1, "p_neq_warm": 5000, "fk_tmax": 0.3, "fk_caps": 96,
+    "ema": 0.999, "aug_pre_drop": 0.15, "aug_pre_sigma": 0.1, "aug_dq_sigma": 0.3,
+    "ckpt_min": 10, "eval_every": 2000, "eval_tasks": 512, "log_every": 50, "patience_evals": 5
+  }
+}

ml/configs/distill.json

@@ -0,0 +1,41 @@
+{
+  "model": {
+    "d": 192,
+    "heads": 8,
+    "ff": 768,
+    "enc": 4,
+    "dec": 4
+  },
+  "train": {
+    "batch": 1024,
+    "lr": 0.0001,
+    "warmup": 1000,
+    "steps": 3000,
+    "wd": 0.01,
+    "betas": [
+      0.9,
+      0.95
+    ],
+    "amp": "bf16",
+    "grad_clip": 1.0,
+    "w_mf": 1.0,
+    "w_fk": 0.1,
+    "w_smooth": 0.01,
+    "w_value": 0.0,
+    "w_sub": 0.0,
+    "p_neq": 0.5,
+    "p_edge": 0.1,
+    "p_neq_warm": 0,
+    "fk_tmax": 0.3,
+    "fk_caps": 96,
+    "ema": 0.999,
+    "aug_pre_drop": 0.15,
+    "aug_pre_sigma": 0.1,
+    "aug_dq_sigma": 0.3,
+    "ckpt_min": 10,
+    "eval_every": 1000,
+    "eval_tasks": 256,
+    "log_every": 50,
+    "patience_evals": 3
+  }
+}

ml/configs/small.json

@@ -1,10 +1,11 @@
 {
   "model": {"d": 128, "heads": 8, "ff": 512, "enc": 4, "dec": 4},
   "train": {
-    "batch": 1024, "lr": 1e-3, "warmup": 1000, "steps": 60000, "wd": 0.01, "betas": [0.9, 0.95],
+    "batch": 1024, "lr": 1e-3, "warmup": 1000, "steps": 24000, "wd": 0.01, "betas": [0.9, 0.95],
     "amp": "bf16", "grad_clip": 1.0,
     "w_mf": 1.0, "w_fk": 0.5, "w_smooth": 0.05, "w_value": 0.1, "w_sub": 0.1,
-    "p_neq": 0.25, "p_neq_warm": 5000, "fk_tmax": 0.3, "fk_caps": 96,
+    "p_neq": 0.5, "p_edge": 0.1, "p_neq_warm": 5000, "fk_tmax": 0.3, "fk_caps": 96,
+    "ema": 0.999, "aug_pre_drop": 0.15, "aug_pre_sigma": 0.1, "aug_dq_sigma": 0.3,
     "ckpt_min": 10, "eval_every": 2000, "eval_tasks": 512, "log_every": 50, "patience_evals": 5
   }
 }

ml/configs/wta.json

@@ -1,10 +1,11 @@
 {
   "model": {"d": 192, "heads": 8, "ff": 768, "enc": 4, "dec": 4, "head": "wta", "K": 8},
   "train": {
-    "batch": 1024, "lr": 1e-3, "warmup": 1000, "steps": 60000, "wd": 0.01, "betas": [0.9, 0.95],
+    "batch": 1024, "lr": 1e-3, "warmup": 1000, "steps": 24000, "wd": 0.01, "betas": [0.9, 0.95],
     "amp": "bf16", "grad_clip": 1.0,
     "w_mf": 1.0, "w_fk": 0.5, "w_smooth": 0.05, "w_value": 0.1, "w_sub": 0.1,
-    "p_neq": 0.25, "p_neq_warm": 5000, "fk_tmax": 0.3, "fk_caps": 96,
+    "p_neq": 0.5, "p_edge": 0.1, "p_neq_warm": 5000, "fk_tmax": 0.3, "fk_caps": 96,
+    "ema": 0.999, "aug_pre_drop": 0.15, "aug_pre_sigma": 0.1, "aug_dq_sigma": 0.3,
     "ckpt_min": 10, "eval_every": 2000, "eval_tasks": 512, "log_every": 50, "patience_evals": 5
   }
 }

ml/dagger/collect.py

@@ -5,7 +5,9 @@
   - 撞了的题：首次碰撞前 0.2 s 的状态（策略"快要撞上"的地方，最有用）；
   - 另取一个 [0.3 s, 首撞前或结束] 内均匀随机时刻的状态（偏离老师的一般状态）。
 起点必须无碰且间隙 > MIN_GAP_MM（老师从碰撞态出发规划不出来，与 build_dataset 的增广接受同一地板），且离目标 > 5 mm（否则 DLS 接管，不需要老师）。
-输出 tasks_XX.jsonl（按 --parts 均分，供 gen_dataset.cjs --tasks 并行），每行 {q0, T, R, cat, src}。
+输出 tasks_XX.jsonl（按 --parts 均分，供 gen_dataset.cjs --tasks 并行），每行 {q0, qPrev, T, R, cat, src}。
+2026-10-19（审查 F7 / C10）：同时记录策略在该时刻的上一帧 qPrev；老师重规划的时间律从当前速度（v0）起步，
+  t=0 样本的 dq 特征用真实上一帧——原来一律按静止起步标注，DAgger 样本不是策略真正到过的 (q, dq) 状态。
 为什么需要：纯行为克隆在闭环里会偏出老师分布（首次 21850 步闭环到达 0%），原方案指望 APG/挖错修，但它们没法从 0% 起步。
 老师只用于训练标注，部署仍然无老师。"""
 import argparse, json, os, random
@@ -26,20 +28,25 @@ def main():
     ap.add_argument('--parts', type=int, default=8); ap.add_argument('--out', required=True)
     ap.add_argument('--bs', type=int, default=64)
     ap.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu')
+    ap.add_argument('--fb_gain', type=float, default=0.0, help='部署分布采集：目标反馈增益（与 rollout 一致）')
+    ap.add_argument('--safety', type=int, default=0, help='部署分布采集：是否开安全层')
+    ap.add_argument('--replan', default='r3'); ap.add_argument('--commit_n', type=int, default=3)
+    ap.add_argument('--safety_partial', type=int, default=0)
     a = ap.parse_args()
     gap_min = float(os.environ.get('MIN_GAP_MM') or 3)
     os.makedirs(a.out, exist_ok=True)
     kin = Kin(device=a.device)
-    ck = torch.load(a.ckpt, map_location=a.device)
+    from ml.common.ckpt import load_policy      # 2026-10-19：EMA 权重 + act_scale
+    model, ck = load_policy(a.ckpt, a.device)
     assert ck.get('fp') in (None, kin.fp), '碰撞模型指纹与 checkpoint 不符'
-    model = build(ck['cfg']).to(a.device); model.load_state_dict(ck['model']); model.eval()
     pool = [t for t in load_tasks(a.samples[:1]) if t['split'] == 'train' and t.get('cat') not in ('D',)]
     rng = random.Random(a.seed); rng.shuffle(pool); tasks = pool[:a.n]
     out, stat = [], {'tasks': len(tasks), 'collided': 0, 'states': 0, 'rej_gap': 0, 'rej_near': 0}
     for i in range(0, len(tasks), a.bs):
         tb = tasks[i:i + a.bs]
         with torch.no_grad():
-            _, per = rollout(model, kin, tb, a.device, K=8, safety=False, seed=a.seed * 100003 + i, return_traj=True)
+            _, per = rollout(model, kin, tb, a.device, K=8, safety=bool(a.safety), seed=a.seed * 100003 + i, return_traj=True,
+                            fb_gain=a.fb_gain, replan=a.replan, commit_n=a.commit_n, safety_partial=a.safety_partial)
         tr = per['traj'].to(a.device)                                   # [B,T,6]
         B, Tn, _ = tr.shape
         _, T, R, _, _ = tasks_to_tensors(tb, a.device)
@@ -61,10 +68,10 @@ def main():
             if c - 1 > lo: cand.append(rng.randint(lo, c - 1))
             for k in cand:
                 if g[b, k] <= gap_min: stat['rej_gap'] += 1; continue
-                q = tr[b, k:k + 1]
+                q = tr[b, k:k + 1]; qprev = tr[b, max(0, k - 1)]
                 e, oe = goal_err(kin, q, T[b:b + 1], R[b:b + 1])
                 if e.item() < 5 and (not t.get('R') or oe.item() < 5): stat['rej_near'] += 1; continue
-                out.append({'q0': [round(v, 3) for v in q[0].tolist()], 'T': t['T'], 'R': t.get('R'), 'cat': t.get('cat', 'A'),
+                out.append({'q0': [round(v, 3) for v in q[0].tolist()], 'qPrev': [round(v, 3) for v in qprev.tolist()], 'T': t['T'], 'R': t.get('R'), 'cat': t.get('cat', 'A'),
                             'src': {'tid': t['tid'], 'k': k, 'col': bool(hit.numel())}})
         del tr, g, per
         if a.device.startswith('cuda'): torch.cuda.empty_cache()

ml/dagger/distill_fb.py

@@ -0,0 +1,57 @@
+#!/usr/bin/env python
+"""反馈蒸馏：用最终部署配置（目标反馈 + 安全层 + 部分前缀 + 每步规划）跑闭环，
+收集 (特征, 复合动作 = 模型输出 + 目标反馈) 样本，供 train.py 微调把反馈内化进策略。
+  PYTHONUTF8=1 python -m ml.dagger.distill_fb --ckpt ml/runs/base_dag2_dag1/best.pt --n 500 \
+      --fb_gain 1.8 --out ml/data/samples_fbdistill
+输出可直接被 ml.data.Samples 读取（collect.X/Y/V/SG/TID.npy；无 meta = act_scale 1.0，与主数据一致）。
+动作标签 = 候选动作（含反馈）的归一化 act（/STEP，clamp ±1），与主数据同口径。
+"""
+import argparse, os, random
+import numpy as np
+import torch
+
+from ml.common.ckpt import load_policy
+from ml.common.features import STEP
+from ml.common.kin_torch import Kin
+from ml.data import load_tasks
+from ml.sim.closed_loop import rollout
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument('--ckpt', required=True); ap.add_argument('--n', type=int, default=500)
+    ap.add_argument('--bs', type=int, default=64); ap.add_argument('--fb_gain', type=float, default=1.8)
+    ap.add_argument('--samples', default='ml/data/samples'); ap.add_argument('--out', required=True)
+    ap.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu'); ap.add_argument('--seed', type=int, default=1)
+    a = ap.parse_args()
+    dev = a.device
+    model, _ = load_policy(a.ckpt, dev)
+    kin = Kin(device=dev)
+    pool = [t for t in load_tasks([a.samples]) if t['split'] == 'train' and t.get('cat') not in ('D',)]
+    random.Random(a.seed).shuffle(pool); tasks = pool[:a.n]
+    stp = torch.tensor(STEP)   # collect 返回 CPU 张量，标签换算也在 CPU
+    X, Y, TID = [], [], []
+    for i in range(0, len(tasks), a.bs):
+        ch = tasks[i:i + a.bs]
+        coll = []
+        with torch.no_grad():
+            rollout(model, kin, ch, dev, K=8, safety=True, replan='empty', fb_gain=a.fb_gain,
+                    safety_partial=1, commit_n=1, seed=a.seed * 100003 + i, collect=coll)
+        for f, ac in coll:
+            X.append(f.numpy()); Y.append((ac / stp).clamp(-1, 1).numpy())
+        ncol = sum(f.shape[0] for f, _ in coll)
+        TID.append(np.arange(ncol, dtype=np.int32) + i * 100000)
+        print(f'[distill] {i + len(ch)}/{len(tasks)} tasks, samples {sum(x.shape[0] for x in X)}', flush=True)
+    os.makedirs(a.out, exist_ok=True)
+    X = np.concatenate(X); Y = np.concatenate(Y); TID = np.concatenate(TID)
+    n = X.shape[0]
+    np.save(os.path.join(a.out, 'collect.X.npy'), X.astype(np.float32))
+    np.save(os.path.join(a.out, 'collect.Y.npy'), Y.astype(np.float32))
+    np.save(os.path.join(a.out, 'collect.V.npy'), np.full(n, 0.5, np.float32))
+    np.save(os.path.join(a.out, 'collect.SG.npy'), np.zeros((n, 13), np.float32))
+    np.save(os.path.join(a.out, 'collect.TID.npy'), TID)
+    print(f'done: {n} samples, X {X.shape}, Y {Y.shape}', flush=True)
+
+
+if __name__ == '__main__':
+    main()

ml/dagger/run.sh

@@ -8,6 +8,9 @@
 #         最终模型路径写进 ml/runs/<tag>_dagger.ckpt（m5 读它）。
 #   断点续跑：采集以 collect.json 为完成标记；老师 shard 先写 .part 再改名、未写完的从断点续写；打包以 .tasks.jsonl 为完成标记；
 #            微调用该轮 last.pt 续训；整轮完成打 ml/runs/<tag>_dag<k>/done。
+# 2026-10-19（审查 review_model_train P0-2）：每轮起点 = python -m ml.common.ckpt pick <目录>：best.pt 的闭环 success > 0 才用它，
+#   否则用 last.pt（success 全 0 时 best.pt 只由开环误差决定、可能落在 p_neq 预热期）；train.py --init 载入 EMA 权重，
+#   只有起点已走过自己的 p_neq 预热时才不再预热。微调的 val 只取主数据（DAgger 样本全部算 train）。
 set -e -o pipefail
 cd "$(dirname "$0")/../.."
 PY=${PYTHON:-python}
@@ -19,9 +22,9 @@ SAMPLES="ml/data/samples"
 for ((r=1; r<=R; r++)); do
   D=ml/data/dag_${TAG}_r$r; SD=ml/data/samples_dag_${TAG}_r$r; OUT=ml/runs/${TAG}_dag$r
   SAMPLES="$SAMPLES $SD"
-  if [ -f $OUT/done ]; then echo "[dagger] 第 $r 轮已完成，跳过"; CK=$OUT/best.pt; continue; fi
+  if [ -f $OUT/done ]; then echo "[dagger] 第 $r 轮已完成，跳过"; CK=$($PY -m ml.common.ckpt pick $OUT); continue; fi
   if [ -f $D/collect.json ]; then echo "[dagger] 第 $r 轮起点已采集，跳过"
-  else echo "[dagger] 第 $r 轮：用 $CK 闭环采起点（$N 题）"; $PY -m ml.dagger.collect --ckpt $CK --samples ml/data/samples --n $N --seed $r --parts $P --out $D; fi
+  else echo "[dagger] 第 $r 轮：用 $CK 闭环采起点（$N 题）"; $PY -m ml.dagger.collect --ckpt $CK --samples ml/data/samples --n $N --seed $r --parts $P --out $D --fb_gain ${FB_GAIN:-0} --safety ${DAG_SAFETY:-0} --replan ${DAG_REPLAN:-r3} --commit_n ${DAG_COMMIT_N:-3} --safety_partial ${DAG_PARTIAL:-0}; fi
   cat $D/collect.json
   pids=()
   for f in $D/tasks_*.jsonl; do
@@ -49,13 +52,14 @@ for ((r=1; r<=R; r++)); do
     done
   fi
   mkdir -p $SD
-  todo=(); for sh in $D/shard_dag_*.jsonl; do [ -e "$sh" ] || continue; b=$(basename $sh .jsonl); [ -f $SD/$b.tasks.jsonl ] || todo+=("$sh"); done
+  SIG=$(node ml/prep/build_dataset.cjs --sig)    # 2026-10-19：与主数据同一套打包签名，参数变了重打（见 run_prep.sh）
+  todo=(); for sh in $D/shard_dag_*.jsonl; do [ -e "$sh" ] || continue; b=$(basename $sh .jsonl); grep -qF "\"sig\":\"$SIG\"" $SD/$b.meta.json 2>/dev/null && [ -f $SD/$b.tasks.jsonl ] || todo+=("$sh"); done
   [ ${#todo[@]} = 0 ] || printf '%s\n' "${todo[@]}" | xargs -P "$P" -I{} node ml/prep/build_dataset.cjs {} $SD
   $PY -m ml.data --index $SD | tail -n 3
   res=""; [ -f $OUT/last.pt ] && res="--resume $OUT/last.pt"
   echo "[dagger] 第 $r 轮：微调 $STEPS 步（样本：$SAMPLES）"
   $PY -m ml.train --samples $SAMPLES --out $OUT --config $CFGF --init $CK --steps $STEPS --lr $LR --warmup 500 $res
-  touch $OUT/done; CK=$OUT/best.pt
+  touch $OUT/done; CK=$($PY -m ml.common.ckpt pick $OUT)
 done
 echo "$CK" > ml/runs/${TAG}_dagger.ckpt
 echo "[dagger] 完成：后续阶段使用 $CK"

ml/data.py

@@ -1,12 +1,17 @@
-"""样本加载：ml/data/samples/*.{X,Y,V,SG,TID}.npy + *.tasks.jsonl（ml/prep/build_dataset.cjs 产出）
+"""样本加载：ml/data/samples/*.{X,Y,V,SG,TID}.npy + *.tasks.jsonl + *.meta.json（ml/prep/build_dataset.cjs 产出）
   python -m ml.data --index ml/data/samples      # 汇总样本/任务数，写 index.json
-全部载入内存（200 万样本约 1.5 GB），按任务 split 字段划分 train / val。"""
+全部载入内存（200 万样本约 1.5 GB），按任务 split 字段划分 train / val。
+动作尺度（2026-10-19）：Y = 每步增量 / (act_scale·STEP)。act_scale 取自各 shard 的 meta.json（DATA_VER 2 起写；
+  旧数据没有 meta → act_scale 1.0、data_ver 1，即原口径）。同一次训练的所有目录（主数据 + DAgger）必须同一个 act_scale。"""
 import argparse, glob, json, os
 import numpy as np
 import torch
 
+PREFIX_SL, MASK_SL, DQ_SL, SUB_SL = slice(47, 65), slice(65, 68), slice(15, 21), slice(68, 75)   # 特征分组（features.py 拼接顺序）
+
 
 def load_tasks(dirs):
+    if isinstance(dirs, str): dirs = [dirs]
     tasks = []
     for d in dirs:
         for f in sorted(glob.glob(os.path.join(d, '*.tasks.jsonl'))):
@@ -16,10 +21,28 @@ def load_tasks(dirs):
     return tasks
 
 
+def read_meta(d):
+    """目录内各 shard 的打包参数（data_ver、act_scale、增广参数、sig）。没有 meta 的旧打包 → {'data_ver': 1, 'act_scale': 1.0}。
+    一部分有、一部分没有，或 act_scale 不一致 → 拒绝（新旧标签尺度混在一起会让动作幅度错 25%）。"""
+    xs = sorted(glob.glob(os.path.join(d, '*.X.npy')))
+    ms = [json.load(open(f[:-6] + '.meta.json', encoding='utf-8')) for f in xs if os.path.exists(f[:-6] + '.meta.json')]
+    if not ms: return {'data_ver': 1, 'act_scale': 1.0}
+    if len(ms) != len(xs): raise SystemExit(f'{d}：{len(xs) - len(ms)} 个 shard 缺 meta.json（新旧打包混在一起），重跑 prep')
+    sc = {float(m.get('act_scale', 1.0)) for m in ms}
+    if len(sc) != 1: raise SystemExit(f'{d}：各 shard 的 act_scale 不一致 {sorted(sc)}，重跑 prep')
+    return ms[0]
+
+
 class Samples:
-    def __init__(self, dirs, device='cpu', split=None):
+    def __init__(self, dirs, device='cpu', split=None, val_dirs=None):
+        """val_dirs：只用这些目录的 val 任务划 val（默认 = 全部 dirs）。2026-10-19（审查 review_closedloop F11）：
+        DAgger 目录的任务起点带速度、来自策略状态，不该混进预训练的 val；train.py / apg 传主数据目录，DAgger 样本全部算 train。"""
         if isinstance(dirs, str): dirs = [dirs]
         xs, ys, vs, ss, ts = [], [], [], [], []
+        metas = [read_meta(d) for d in dirs]
+        sc = {float(m['act_scale']) for m in metas}
+        if len(sc) > 1: raise SystemExit(f'样本目录的 act_scale 不一致 {sorted(sc)}（{dirs}）：DAgger 样本须与主数据同一版打包')
+        self.act_scale = sc.pop(); self.data_ver = int(metas[0].get('data_ver', 1)) if metas else 1
         for d in dirs:
             for fx in sorted(glob.glob(os.path.join(d, '*.X.npy'))):
                 b = fx[:-6]
@@ -27,7 +50,7 @@ class Samples:
                 ss.append(np.load(b + '.SG.npy')); ts.append(np.load(b + '.TID.npy'))
         if not xs: raise SystemExit(f'没有样本：{dirs}（先跑 ml/prep/run_prep.sh）')
         self.tasks = load_tasks(dirs)
-        val_tid = {t['tid'] for t in self.tasks if t['split'] == 'val'}
+        val_tid = {t['tid'] for t in load_tasks(val_dirs if val_dirs is not None else dirs) if t['split'] == 'val'}
         X, Y, V, SG, TID = map(np.concatenate, (xs, ys, vs, ss, ts))
         is_val = np.isin(TID, np.fromiter(val_tid, dtype=np.int64)) if val_tid else np.zeros(len(TID), bool)
         sel = {'train': ~is_val, 'val': is_val, None: np.ones(len(TID), bool)}[split]
@@ -35,9 +58,43 @@ class Samples:
         self.X, self.Y, self.V, self.SG, self.TID = t(X), t(Y), t(V), t(SG), t(TID)
         self.n = self.X.shape[0]
 
-    def batch(self, bs, gen=None):
+    def batch(self, bs, gen=None, aug=None):
+        """aug：训练配置（train 段），按 aug_pre_drop / aug_pre_sigma / aug_dq_sigma 做条件输入增广（见 augment）；None = 原样。"""
         i = torch.randint(0, self.n, (bs,), device=self.X.device, generator=gen)
-        return self.X[i], self.Y[i], self.V[i], self.SG[i]
+        X, Y = self.X[i], self.Y[i]
+        if aug: X, Y = self.augment(X.clone(), Y.clone(), aug, gen)
+        return X, Y, self.V[i], self.SG[i]
+
+    def augment(self, X, Y, cfg, gen=None):
+        """条件输入增广（2026-10-19，审查 review_data_actions C3 / review_model_train P1-4：前缀与 dq 的 copycat 捷径）。
+        训练集里 r=3 时要执行的 Y[3:6] 几乎就是前缀第 3 步的延续，dq 又永远是老师的干净速度，模型容易学成「照抄上一块」，
+        闭环里自己的误差经前缀 / dq 原样带回。三种扰动，标签都保持自洽：
+          - 前缀丢弃（aug_pre_drop）：r>0 的行以该概率清空前缀与掩码 → 变成 r=0 样本。Y 本来就从当前 q 起算，r=0 时标签不变；
+          - 前缀噪声（aug_pre_sigma，STEP 为单位）：r>0 的行一半加噪声 ε。前缀 = 已提交、一定会执行的步，所以标签前 r 步同样 +ε，
+            第 r 步减去 Σε 把位置拉回老师轨迹（亏空在下一步补齐，与增广收回同理）；
+          - dq 噪声（aug_dq_sigma，STEP 为单位）：一半的行给速度特征加噪声，状态与标签不变。"""
+        B, dev = X.shape[0], X.device
+        rnd = lambda *s: torch.rand(*s, device=dev, generator=gen)
+        rn = lambda *s: torch.randn(*s, device=dev, generator=gen)
+        pl = X[:, MASK_SL].sum(-1).round().long()
+        pd = float(cfg.get('aug_pre_drop', 0.0))
+        if pd > 0:
+            m = (rnd(B) < pd) & (pl > 0)
+            X[:, 47:68] = torch.where(m[:, None], torch.zeros_like(X[:, 47:68]), X[:, 47:68]); pl = torch.where(m, torch.zeros_like(pl), pl)
+        ps = float(cfg.get('aug_pre_sigma', 0.0))
+        if ps > 0:
+            m = (rnd(B) < 0.5) & (pl > 0)
+            km = (torch.arange(3, device=dev)[None] < pl[:, None]) & m[:, None]                 # [B,3] 前缀有效步
+            eps = rn(B, 3, 6) * ps * km[..., None].to(X.dtype)                                  # STEP 为单位
+            X[:, PREFIX_SL] = X[:, PREFIX_SL] + eps.reshape(B, 18)
+            Y[:, :3] = Y[:, :3] + eps / self.act_scale                                          # Y 以 act_scale·STEP 为单位
+            Y[torch.arange(B, device=dev), pl.clamp(max=Y.shape[1] - 1)] -= eps.sum(1) / self.act_scale
+            Y.clamp_(-1, 1)
+        ds = float(cfg.get('aug_dq_sigma', 0.0))
+        if ds > 0:
+            m = (rnd(B) < 0.5).to(X.dtype)
+            X[:, DQ_SL] = X[:, DQ_SL] + rn(B, 6) * ds * m[:, None]
+        return X, Y
 
 
 def main():
@@ -46,7 +103,9 @@ def main():
     tasks = load_tasks([a.index])
     cats = {}
     for t in tasks: cats[t['cat'] + '/' + t['mode']] = cats.get(t['cat'] + '/' + t['mode'], 0) + 1
-    info = {'train_samples': tr.n, 'val_samples': va.n, 'tasks': len(tasks), 'val_tasks': sum(t['split'] == 'val' for t in tasks), 'by_cat_mode': cats}
+    meta = read_meta(a.index)
+    info = {'train_samples': tr.n, 'val_samples': va.n, 'tasks': len(tasks), 'val_tasks': sum(t['split'] == 'val' for t in tasks), 'by_cat_mode': cats,
+            'data_ver': int(meta.get('data_ver', 1)), 'act_scale': float(meta.get('act_scale', 1.0)), 'sig': meta.get('sig')}
     json.dump(info, open(os.path.join(a.index, 'index.json'), 'w'), indent=1)
     print(json.dumps(info, indent=1))
 

ml/eval/acceptance.py

@@ -4,7 +4,7 @@
 其余项读已有文件（缺文件记为 missing，不算通过）：
   ml/runs/policy_bench.json   部署形态：各集合穿模帧 = 0；near 30/30（pos、pose）；route 全部到达；nc1000 0 失败；刹停率 < 2%；R8 恢复 ≥ 95%、活锁 = 0
   ml/data/r2.json             R2 近邻标签 std < 10% 步长
-  ml/runs/bench_*/log.jsonl   R10 吞吐 ≥ 5000 样本/s（≈ 60000 步 ≤ 3.5 h）
+  ml/runs/bench_*/log.jsonl   R10 吞吐 ≥ 5000 样本/s（24000 步 batch 1024 ≈ 1.4 h）
   ml/runs/m5_gate_*.json      M5 头门禁
   ml/runs/mine/r*/gate.json   R7 最后一轮左右比例 / 行程门禁
   ml/runs/r4.json             R4 浏览器推理 batch=8 P99 < 15 ms（r4 探针）
@@ -37,8 +37,8 @@ def main():
         items.append({'item': name, 'value': value, 'pass': (None if ok is None else bool(ok)), 'rule': rule})
 
     # 1) held-out 闭环（仿真安全层）
-    ck = torch.load(a.ckpt, map_location=a.device)
-    model = build(ck['cfg']).to(a.device); model.load_state_dict(ck['model']); model.eval()
+    from ml.common.ckpt import load_policy      # 2026-10-19：EMA 权重 + act_scale
+    model, ck = load_policy(a.ckpt, a.device)
     kin = Kin(device=a.device)
     val = [t for t in load_tasks([a.samples]) if t['split'] == 'val']
     reach = [t for t in val if t.get('reached') and t['cat'] != 'D']
@@ -97,7 +97,7 @@ def main():
     for f in sorted(glob.glob('ml/runs/bench_*/log.jsonl')):
         recs = [json.loads(l) for l in open(f) if '"samples_per_s"' in l]
         if recs: sps = max(sps or 0, recs[-1]['samples_per_s'])
-    add('R10 训练吞吐（样本/s）', sps, sps is not None and sps >= 5000, '≥ 5000（≈ 60000 步 ≤ 3.5 h）')
+    add('R10 训练吞吐（样本/s）', sps, sps is not None and sps >= 5000, '≥ 5000（24000 步 ≈ 1.4 h）')
     m5 = [jload(f) for f in sorted(glob.glob('ml/runs/m5_gate_*.json'))]
     m5 = [m for m in m5 if m]
     add('M5 头门禁', [{'head': m['head'], 'collision_rate': m['policy_only']['collision_rate']} for m in m5] or 'missing',