#!/usr/bin/env python
"""闭环失败定位（只读，手动运行，约 2–4 分钟；训练进行中也可以另开终端跑，显存约 2–3 GB）。

2026-10-19 新增：第二次预训练（979a581）开环已经正常（候选0 末端误差 ~6 mm、K 中最好 ~2.9 mm、多样性 0.65），
闭环却仍 100% 碰撞、0% 到达。开环用的是 JS 生成的特征 X，闭环用的是 Python features() 现算的特征，
两者之间只要有一处口径不同，模型在闭环里看到的就是没见过的输入。本脚本回答三件事：
  A. 特征口径：val 任务起点（非增广、老师状态）上，Python features() 与数据集 X 的逐组最大差；
  B. 同一噪声下，用两份特征得到的动作差（16 步末端关节差，度）——A 有差但 B 很小说明差异无害；
  C. 闭环失败形态：首次碰撞时刻分布、碰撞步数、首次碰撞的组对、全程最小目标误差；
     以及同一批题开安全层（safety=True，近似部署）的到达率 / 刹停次数。
用法（仓库根目录）：
  PYTHONUTF8=1 python -m ml.eval.diag_cl --ckpt ml/runs/base/last.pt --n 256 --out ml/runs/diag_cl.json
"""
import argparse, json, os

import torch

from ml.common.kin_torch import Kin
from ml.common.features import features, STEP, H
from ml.data import Samples, load_tasks
from ml.model.policy import build
from ml.sim.closed_loop import rollout, goal_err, DT

# 83 维特征分组（与 features.py 拼接顺序一致）
GROUPS = [('sincos', 0, 12), ('joint_norm', 12, 15), ('dq', 15, 21), ('p', 21, 24), ('Rq', 24, 30), ('T', 30, 33),
          ('R', 33, 39), ('T-p', 39, 42), ('T-p_clamp', 42, 45), ('mode', 45, 47), ('prefix', 47, 65), ('mask', 65, 68),
          ('sub', 68, 74), ('sub_flag', 74, 75), ('gaps', 75, 83)]


PROG_T = (0.5, 1.0, 2.0)     # 目标误差 / 起点误差 的取样时刻（s）


def pct(v, p):
    v = sorted(v)
    return v[min(len(v) - 1, int(len(v) * p))] if v else None


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument('--ckpt', required=True); ap.add_argument('--samples', default='ml/data/samples')
    ap.add_argument('--n', type=int, default=256); ap.add_argument('--bs', type=int, default=64)
    ap.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu'); ap.add_argument('--out', default='ml/runs/diag_cl.json')
    a = ap.parse_args()
    dev = a.device
    ck = torch.load(a.ckpt, map_location=dev)
    model = build(ck['cfg']).to(dev); model.load_state_dict(ck['model']); model.eval()
    kin = Kin(device=dev)
    out = {'ckpt': a.ckpt, 'step': ck.get('step')}
    tasks = [t for t in load_tasks([a.samples]) if t['split'] == 'val'][:a.n]
    by_tid = {t['tid']: t for t in tasks}

    # ---------- A / B：起点特征口径 ----------
    vd = Samples(a.samples, device=dev, split='val')
    tid = vd.TID.long()
    first = torch.ones_like(tid, dtype=torch.bool); first[1:] = tid[1:] != tid[:-1]     # 每题第一行 = t=0 样本
    idx = [i for i in first.nonzero().flatten().tolist() if int(tid[i]) in by_tid]
    X0 = vd.X[idx].float()
    tk = [by_tid[int(tid[i])] for i in idx]
    q0 = torch.tensor([t['q0'] for t in tk], dtype=torch.float32, device=dev)
    qx = torch.atan2(X0[:, 0:12:2], X0[:, 1:12:2]) * 180 / torch.pi
    dqx = torch.remainder(qx - q0 + 180, 360) - 180
    clean = dqx.abs().max(-1).values < 1e-2                  # 未被噪声增广的起点
    out['A_rows'] = int(clean.sum()); out['A_rows_aug_skipped'] = int((~clean).sum())
    if clean.any():
        Xc = X0[clean]; qc = q0[clean]; tc = [t for t, c in zip(tk, clean.tolist()) if c]
        T = torch.tensor([t['T'] for t in tc], dtype=torch.float32, device=dev)
        R = torch.tensor([t['R'] if t.get('R') else [0.] * 9 for t in tc], dtype=torch.float32, device=dev)
        B = qc.shape[0]
        pre = Xc[:, 47:65].view(B, 3, 6) * torch.tensor(STEP, device=dev)       # 前缀 / 子目标直接取数据集里的，只比状态相关部分
        pl = Xc[:, 65:68].sum(-1).round()
        sf = Xc[:, 74]
        sub = qc + Xc[:, 68:74] * 180
        Fp = features(kin, qc, qc, T, R, pre, pl, sub, sf)
        out['A_maxdiff'] = {g: (Fp[:, s:e] - Xc[:, s:e]).abs().max().item() for g, s, e in GROUPS}
        out['A_meandiff'] = {g: (Fp[:, s:e] - Xc[:, s:e]).abs().mean().item() for g, s, e in GROUPS}
        with torch.no_grad():
            z = torch.randn(B * 8, H, 6, device=dev, generator=torch.Generator(device=dev).manual_seed(7))
            a1, _, _ = model.sample_k(Xc, 8, z); a2, _, _ = model.sample_k(Fp, 8, z)
        stepv = torch.tensor(STEP, device=dev)
        d = ((a1.clamp(-1, 1) - a2.clamp(-1, 1)) * stepv).sum(2).abs().max(-1).values     # [B,K] 16 步末端关节差（度）
        mag = (a1.clamp(-1, 1) * stepv).sum(2).abs().max(-1).values
        out['B_act_diff_deg'] = {'mean': d.mean().item(), 'p90': d.flatten().kthvalue(max(1, int(0.9 * d.numel()))).values.item(),
                                 'ref_motion_deg_mean': mag.mean().item()}

    # ---------- C：闭环失败形态 ----------
    first_t, col_steps, pairs, min_err, reach = [], [], {}, [], 0
    prog = {ts: [] for ts in PROG_T}
    reach_s, brakes_s, col_s = 0, [], 0
    for i in range(0, len(tasks), a.bs):
        ch = tasks[i:i + a.bs]
        with torch.no_grad():
            _, per = rollout(model, kin, ch, dev, K=8, safety=False, seed=i, return_traj=True)
            tr = per['traj'].to(dev)                                            # [B,Tn,6]
            B, Tn, _ = tr.shape
            flat = tr.reshape(-1, 6)                                            # 分块 ≤1024 点（同 dagger/collect.py 的显存修复）
            gm = torch.cat([kin.gaps(flat[s:s + 1024], 0.5)['gm'] for s in range(0, flat.shape[0], 1024)]).view(B, Tn, 49)
            hit = gm.min(-1).values <= 0                                        # [B,Tn]
            Tt = torch.tensor([t['T'] for t in ch], dtype=torch.float32, device=dev)
            Rt = torch.tensor([t['R'] if t.get('R') else [0.] * 9 for t in ch], dtype=torch.float32, device=dev)
            e, _ = goal_err(kin, tr.reshape(-1, 6), Tt.repeat_interleave(Tn, 0), Rt.repeat_interleave(Tn, 0))
            e = e.view(B, Tn)
            for b in range(B):
                for ts in PROG_T:
                    k = min(Tn - 1, int(round(ts / DT))); prog[ts].append((e[b, k] / e[b, 0].clamp(min=1e-3)).item())
            # 到达后轨迹被冻结（d=0），之后不再计
            for b in range(B):
                if per['done'][b]: reach += 1
                min_err.append(e[b].min().item()); col_steps.append(int(per['col'][b]))
                h = hit[b].nonzero().flatten()
                if len(h):
                    k = int(h[0]); first_t.append(k * DT)
                    pr = int(gm[b, k].argmin()); key = '%d-%d' % (pr // 7, pr % 7); pairs[key] = pairs.get(key, 0) + 1
            rs, ps = rollout(model, kin, ch, dev, K=8, safety=True, seed=i)
            reach_s += int(ps['done'].sum()); col_s += int((ps['col'] > 0).sum()); brakes_s += ps['brakes'].tolist()
    n = len(tasks)
    out['C'] = {'n': n, 'reach': reach / n, 'collided': len(first_t) / n,
                'first_hit_s': {'p10': pct(first_t, .1), 'p50': pct(first_t, .5), 'p90': pct(first_t, .9),
                                'within_0.1s': sum(t <= 0.1 for t in first_t) / max(1, len(first_t))},
                'col_steps': {'p50': pct(col_steps, .5), 'p90': pct(col_steps, .9)},
                'first_hit_pair(group a-b)': dict(sorted(pairs.items(), key=lambda kv: -kv[1])[:8]),
                'min_goal_err_mm': {'p10': pct(min_err, .1), 'p50': pct(min_err, .5), 'p90': pct(min_err, .9)},
                'safety_on': {'reach': reach_s / n, 'collided': col_s / n, 'brakes_p50': pct(brakes_s, .5)}}
    out['C']['goal_err_ratio_p50'] = {'%.1fs' % ts: pct(v, .5) for ts, v in prog.items()}   # <1 = 在靠近目标

    # ---------- D：子目标是不是元凶（2026-10-19 加：C 显示全程最小目标误差中位 254 mm，策略根本没朝目标走） ----------
    # 训练时特征里的子目标是老师真实途经点（80% 概率给），闭环里却换成子目标头自己的预测；预测偏了，策略就会忠实地朝错的方向走。
    sg_by_tid = {}
    for i, t in zip(idx, tk):
        qrow = torch.atan2(vd.X[i, 0:12:2].float(), vd.X[i, 1:12:2].float()) * 180 / torch.pi
        sgr = vd.SG[i].float()
        sg_by_tid[t['tid']] = (qrow[None] + sgr[:12].view(2, 6) * 180, bool(sgr[12] > 0.5))
    # D0：冷启动子目标头精度（起点特征里抹掉数据集的真子目标，与闭环冷启动同口径）
    Xs = X0.clone(); Xs[:, 68:75] = 0
    with torch.no_grad(): _, sp = model.heads(model.encode(Xs))
    lab = torch.stack([vd.SG[i].float() for i in idx])
    need_l = lab[:, 12] > 0.5; need_p = sp[:, 12] > 0.5
    via_err = ((sp[:, :6] - lab[:, :6]) * 180).abs().max(-1).values
    out['D0_subhead'] = {'n': len(idx), 'need_label_rate': need_l.float().mean().item(), 'need_acc': (need_l == need_p).float().mean().item(),
                         'via1_maxjoint_err_deg_p50': pct(via_err[need_l].tolist(), .5), 'via1_maxjoint_err_deg_p90': pct(via_err[need_l].tolist(), .9)}
    # D1 不用子目标、D2 用老师真实途经点（oracle），同一批题、同一随机种子，对比 C 的基线
    def run(tag, **kw):
        rc = cc_ = 0; me = []
        for i in range(0, len(tasks), a.bs):
            ch = tasks[i:i + a.bs]; extra = dict(kw)
            if kw.get('sub_override') == 'oracle':
                subs = torch.stack([sg_by_tid[t['tid']][0] if t['tid'] in sg_by_tid else torch.zeros(2, 6, device=dev) for t in ch])
                need = torch.tensor([sg_by_tid[t['tid']][1] if t['tid'] in sg_by_tid else False for t in ch], device=dev)
                extra['sub_override'] = (subs, need)
            with torch.no_grad():
                _, pp = rollout(model, kin, ch, dev, K=8, safety=False, seed=i, **extra)
            rc += int(pp['done'].sum()); cc_ += int((pp['col'] > 0).sum()); me += pp['err'].tolist()
        out[tag] = {'reach': rc / n, 'collided': cc_ / n, 'end_goal_err_mm_p50': pct(me, .5)}
    run('D1_no_sub', use_sub=False)
    run('D2_oracle_sub', sub_override='oracle')
    s = json.dumps(out, indent=1, ensure_ascii=False); print(s)
    os.makedirs(os.path.dirname(a.out) or '.', exist_ok=True)
    open(a.out, 'w', encoding='utf-8').write(s)


if __name__ == '__main__':
    main()