exp: 轮次39(续)——ep_len 三臂：时限内单调+2.4pp 但行程比单调变差，负结论；改上 ep_waste

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/探索日志.md

@@ -52,6 +52,23 @@
   1. 新加的 `backoff` 指标一度在产物里根本没有——`ml/eval/eval_cl.py:run()` 的 `out` 是**自己显式列键**的，不是从 `rollout` 返回的字典通抄，所以我往 `res` 加的 `backoff_mean` 被无声丢掉。已改成从逐题 `per['backoff']` 汇总（`backoff_mean` + `reject_all_mean = backoff+brakes`，后者才是与"原来全拒次数"可比的量）。**教训：加指标要顺着输出链路走到底，别只改生产端。**
   2. 停后台 GPU 作业时，`TaskStop` 之后 `bash` 脚本链会活下来并继续拉起下一个 python；而 `nvidia-smi --query-compute-apps` 在这台机器上列不出这些 python（只列 ShellHost/wallpaper32 之类），据此判断"卡已空"会误判。我第一次按 PID 杀子进程后，链条 30 秒后又起了一个新的 eval。第二次我按进程树杀，但**把父子关系读反了**（Windows 的 PPID 在父进程已死时是陈旧值），结果误杀了新链的第一臂。正确做法：重启前用命令行正则列出全量候选、数 python 个数确认单租户，重启后再数一次；杀之前先按 `StartTime` 判新旧。backoff 那一臂因此要补跑。
 
+## 行程惩罚 ep_len 三臂读数：时限变好，但行程比没降（轮次 39 续，负结论）
+
+- 同一 256 题、同 seed、都带 `--inc_cap 1`，只改 `select=ep_len --len_w`：
+
+  | 臂 | 到达 | 时限内(≤1.3×老师) | 行程比(中位) | 刹车/题 | 成功题末态误差 |
+  |---|---|---|---|---|---|
+  | 对照（endpoint） | 0.5234 | 0.2656 | 2.6158 | 19.55 | 0.218 mm |
+  | len_w=0.1 | 0.5117 | 0.2695 | 2.6216 | 19.49 | 0.244 mm |
+  | len_w=0.3 | 0.5195 | 0.2852 | 2.6548 | 20.00 | 0.219 mm |
+  | len_w=1.0 | 0.5273 | **0.2891** | 2.6719 | 19.35 | — |
+
+  （`docs/experiment_round39_lenw{0.1,0.3,1.0}.json`）
+- **两条相反的结论**：① 时限内**单调变好**（0.2656→0.2891，+2.4 pp = 6 题），到达在噪声内（±0.4 pp）；② 但**行程比单调变差**（2.6158→2.6719）。
+  ⇒ "候选地平线内的累计行程"并不是"实际执行行程"好的代理：罚它选出来的候选在**近端少过冲**（所以更快收住、时限内变好），但整段任务要更多次规划才能走完，累计行程反而涨。这条是我原先的假设被打掉，不是被确认。
+- 顺着这个读数，真正该罚的是**同一轴来回**而不是总长：已实现 `select='ep_waste'`，罚 `(1 − 同轴净位移/累计行程) × 当前误差`（乘当前误差是为了近端不压死进度项——上一版的形状问题正在这里）。三臂 `len_w=0.1/0.3/1.0` 排在 `ml/runs/_waste_sweep.sh`。
+- 换向率指标 `reverse_p50` 从这一批之后才会出现在产物里（前面的臂是在改动前加载的模块跑的），口径已与离线分解同分母，预期对照臂应复现 **23.5%** 量级——**对不上就要查口径，不能各自报各自的**。
+
 ## 换向率进了评估指标（轮次 39 续）
 
 - 分解既然指到"来回抖"，每个臂都要直接报这个数，不能每次回头读转储：`rollout` 里按轴记方向，`reverse_rate` = 同轴方向翻转的步数 / 活动步数（站住的步不重置方向，所以"停了再同向走"不算换向）；`eval_cl` 报 `reverse_p50`/`reverse_mean`。