"""Generic stage 0 for a local LeRobot v3.0 dataset -> teleop_std_poc-contract npz per episode + index.json with calibration.
Layouts: 'aloha14' (12 joints + grippers at 6,13), 'so100' (5 joints + gripper 5).
    python prep_generic.py --root datasets/aloha_sim_transfer_cube_human --out trajs_aloha_tc --layout aloha14 [--deg]"""
import argparse, glob, json
from pathlib import Path
import numpy as np, pyarrow.parquet as pq

def split(S, layout):
    if layout == "aloha14": return np.concatenate([S[:, 0:6], S[:, 7:13]], 1), np.maximum(S[:, 6], S[:, 13])
    if layout == "so100": return S[:, :5], S[:, 5]
    raise ValueError(layout)

if __name__ == "__main__":
    ap = argparse.ArgumentParser(); ap.add_argument("--root", required=True); ap.add_argument("--out", required=True); ap.add_argument("--layout", default="aloha14"); ap.add_argument("--deg", action="store_true")
    a = ap.parse_args(); out = Path(a.out); out.mkdir(parents=True, exist_ok=True); K = 180 / np.pi if a.deg else 1.0
    tt = pq.read_table(f"{a.root}/meta/tasks.parquet").to_pandas(); tmap = {i: t for t, i in zip(tt.index, tt["task_index"])}
    steps, accs, spans, index = [], [], [], {"source": a.root, "layout": a.layout, "units": "deg" if a.deg else "raw", "episodes": {}}
    for p in sorted(glob.glob(f"{a.root}/data/**/*.parquet", recursive=True)):
        t = pq.read_table(p, columns=["episode_index", "task_index", "observation.state", "action", "timestamp"]).to_pandas()
        for ep, df in t.groupby("episode_index"):
            S = np.stack(df["observation.state"].values).astype(np.float32) * K; A = np.stack(df["action"].values).astype(np.float32) * K; ts = df["timestamp"].values.astype(float)
            P, G = split(S, a.layout); n = len(S)
            np.savez(out / f"ep{int(ep)}.npz", positions=P, gripper=G, actions=A, timestamps=ts, pos=np.arange(n, dtype=float), state=S)
            d = np.linalg.norm(np.diff(P, axis=0), axis=1); steps += list(d[d > 0.2 * np.median(d[d > 0]) if (d > 0).any() else d]); dt = np.median(np.diff(ts))
            v = np.diff(P, axis=0) / dt; accs += list(np.linalg.norm(np.diff(v, axis=0) / dt, axis=1)); spans.append(G.max() - G.min())
            index["episodes"][f"ep{int(ep)}"] = {"source": "human", "frames_src": n, "frames": n, "task": tmap[int(df["task_index"].iloc[0])]}
    index["task"] = next(iter(index["episodes"].values()))["task"]; med, macc, span = float(np.median(steps)), float(np.mean(accs)), float(np.median(spans))
    index["calib"] = {"median_step": med, "mean_acc": macc, "gripper_span": span, "lambda_acc_20pct": 0.25 * float(np.sum(steps)) / float(np.sum(accs)), "gripper_thr_5pct": 0.05 * span, "d_target_k3": 3 * med, "d_target_k6": 6 * med}
    index["n_episodes"] = len(index["episodes"]); (out / "index.json").write_text(json.dumps(index, indent=1)); print(json.dumps({k: index[k] for k in ("n_episodes", "task", "calib")}, indent=1))