"""
ingest_cpcb.py -> data\ingest\   (run: python data\ingest\ingest_cpcb.py)

CPCB ground-truth stations -> grid-aligned SUPERVISION LABELS.
Each station (lat,lon) is dropped into its 10 km cell (containment); stations
in the same cell on the same day are averaged. Output:
    cpcb_labels_daily.parquet : node_id | date | pm2_5 pm10 no2 so2 co o3 nh3 | n_stations | centroid_lat/lon
and sets grid_nodes.has_cpcb_station (the "directly_verified" calibration flag).

Source: AtmosIQ/cpcb/merged_cpcb_clean.csv (HF), or pass --csv <path>.
"""
from __future__ import annotations
import os, sys, argparse, pathlib
os.environ.setdefault("HF_HUB_DISABLE_XET", "1")
sys.path.insert(0, str(pathlib.Path(__file__).resolve().parents[1] / "grid"))
import numpy as np, pandas as pd
from grid_index import GridIndex

OUT  = pathlib.Path(os.environ.get("ATMOS_OUT", "data/ingest"))
GRID = pathlib.Path(os.environ.get("ATMOS_GRID", "data/grid/grid_nodes.parquet"))

# CPCB raw col -> model target name (blueprint P=7). Extras (no, nox_ppb) dropped.
TARGETS = {"pm25":"pm2_5","pm10":"pm10","no2":"no2","so2":"so2","co":"co","o3":"o3","nh3":"nh3"}
# physical caps (CPCB is ground truth; light safety only — drops sensor spikes)
CAP = {"pm2_5":(0,1000),"pm10":(0,1000),"no2":(0,500),"so2":(0,300),
       "co":(0,50),"o3":(0,600),"nh3":(0,600)}

def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--csv", default=None, help="local merged_cpcb_clean.csv (else HF download)")
    args = ap.parse_args()

    gi = GridIndex.from_parquet(str(GRID))
    print(f"grid: {len(gi)} cells")

    if args.csv:
        path = args.csv
    else:
        from huggingface_hub import hf_hub_download
        print("downloading CPCB csv from HF...")
        path = hf_hub_download("AtmosIQ/cpcb", filename="merged_cpcb_clean.csv",
                               repo_type="dataset", token=os.environ.get("HF_TOKEN"))
    print(f"reading {path}")
    df = pd.read_csv(path, usecols=["date","lat","lon"] + list(TARGETS.keys()))
    print(f"raw rows: {len(df):,}")

    # normalise date to YYYY-MM-DD (already IST calendar date)
    df["date"] = pd.to_datetime(df["date"], errors="coerce").dt.strftime("%Y-%m-%d")
    df = df.dropna(subset=["date","lat","lon"])
    df = df.rename(columns=TARGETS)

    # station -> cell (containment)
    df["node_id"] = gi.assign_node_id(df["lat"].values, df["lon"].values)
    dropped = int(df["node_id"].isna().sum())
    df = df[df["node_id"].notna()]
    print(f"stations mapped to grid; {dropped:,} station-days fell outside grid (dropped)")

    # safety clip each target (leaves NaN as NaN)
    for c,(lo,hi) in CAP.items():
        if c in df: df[c] = df[c].clip(lo, hi)

    # aggregate stations sharing a cell/day -> mean target; count stations
    g = df.groupby(["node_id","date"])
    lab = g[list(TARGETS.values())].mean()
    lab["n_stations"] = g["lat"].size()   # station-days averaged (proxy for station count)
    lab = lab.reset_index().merge(gi.centroids(), on="node_id", how="left")

    lab.to_parquet(OUT/"cpcb_labels_daily.parquet", index=False)
    print(f"-> {OUT/'cpcb_labels_daily.parquet'}  ({len(lab):,} labeled node-days)")
    print(f"   unique labeled cells: {lab['node_id'].nunique():,}/{len(gi)} "
          f"| date range {lab['date'].min()}..{lab['date'].max()}")
    for c in TARGETS.values():
        s = lab[c]; print(f"   {c}: fill={100*s.notna().mean():.1f}%  mean={s.mean():.2f}")

    # calibration flag: which cells ever host a station
    gn = pd.read_parquet(GRID)
    stationed = set(lab["node_id"].unique())
    gn["has_cpcb_station"] = gn["node_id"].isin(stationed)
    gn.to_parquet(GRID, index=False)
    print(f"   grid_nodes.has_cpcb_station set: {gn['has_cpcb_station'].sum():,} cells verified")

if __name__ == "__main__":
    main()