r"""add_ystats.py — compute per-pollutant log-target mean/std (TRAIN fold) from existing
tensors and add to norm.npz. Avoids re-running full assemble. Run once before retraining.
Run: python data\models\add_ystats.py --tensors data\tensors
"""
import argparse, pathlib, numpy as np
ap=argparse.ArgumentParser(); ap.add_argument("--tensors",default="data/tensors"); a=ap.parse_args()
t=pathlib.Path(a.tensors)
Y=np.load(t/"targets.npy").astype(np.float32)            # [D,N,7]
M=np.load(t/"target_mask.npy")                           # [D,N,7]
split=np.load(t/"node_split.npy")                        # [N]
tr=(split==0)
yl=np.log(np.clip(Y,0,None)+1e-3)
ylm=np.zeros(7,np.float32); yls=np.ones(7,np.float32)
POLL=["pm2_5","pm10","no2","so2","co","o3","nh3"]
for j in range(7):
    v=yl[:,tr,j][M[:,tr,j]>0]
    if len(v): ylm[j]=v.mean(); yls[j]=v.std()+1e-6
    print(f"  {POLL[j]:6}: log-mean={ylm[j]:+.2f} log-std={yls[j]:.2f}")
nz=dict(np.load(t/"norm.npz"))
nz["y_logmean"]=ylm; nz["y_logstd"]=yls
np.savez(t/"norm.npz",**nz)
print(f"-> updated {t/'norm.npz'} with y_logmean/y_logstd")