"""Rebuild the Game 4 score timeline from play-by-play and compute a win-probability path.
Model: final margin ~ Normal(current margin + edge * t/48, sigma * sqrt(t/48)) (Brownian-motion model).
Two versions: sigma = 13.5 (base) and sigma calibrated so P(NYK win) = 0.4% at 9:33 Q4 (95-75),
matching the published model reading.
Usage: python3 winprob_g4.py shots.csv freethrows.csv out.csv
"""
import sys
import numpy as np
import pandas as pd
from scipy.stats import norm
from scipy.optimize import brentq

shots = pd.read_csv(sys.argv[1])
fts = pd.read_csv(sys.argv[2])
shots = shots[shots.result == "make"].assign(pts=lambda x: np.where(x.shot_type == "3pt", 3, 2), kind=0)
fts = fts[fts.result == "make"].assign(pts=1, kind=1)
ev = pd.concat([shots[["quarter", "clock", "team", "player", "pts", "kind"]],
                fts[["quarter", "clock", "team", "player", "pts", "kind"]]])
secs = ev.clock.str.split(":").apply(lambda x: int(x[0]) * 60 + float(x[1]))
ev["elapsed_min"] = ((ev.quarter - 1) * 720 + (720 - secs)) / 60
ev = ev.sort_values(["elapsed_min", "kind"]).reset_index(drop=True)
ev["nyk"] = np.where(ev.team == "NYK", ev.pts, 0).cumsum()
ev["sas"] = np.where(ev.team == "SAS", ev.pts, 0).cumsum()
ev["margin"] = ev.nyk - ev.sas
ev["t_left"] = 48 - ev.elapsed_min


def wp(margin, t_left, sigma, edge=0.0):
    t = np.maximum(t_left, 1e-6)
    return norm.sf(0, loc=margin + edge * t / 48, scale=sigma * np.sqrt(t / 48))


# --- validation against reported checkpoints ---
def score_at(elapsed):
    s = ev[ev.elapsed_min <= elapsed + 1e-9]
    return (int(s.sas.iloc[-1]), int(s.nyk.iloc[-1])) if len(s) else (0, 0)


checks = {"End Q1 (41-22)": 12, "Half (76-49)": 24, "9:40 Q3 (81-52)": 26 + 20 / 60,
          "End Q3 (90-75)": 36, "9:33 Q4 (95-75)": 38 + 27 / 60, "Final (106-107)": 48}
for k, v in checks.items():
    print(f"{k:18s} -> SAS-NYK {score_at(v)}")
low = ev.loc[ev.margin.idxmin()]
print(f"Low point: NYK {low.margin} at {low.elapsed_min:.2f} min ({low.sas}-{low.nyk})")

# --- calibrate sigma to the published 0.4% at 9:33 Q4, down 20 ---
t933 = 9 + 33 / 60
sig_cal = brentq(lambda s: wp(-20, t933, s) - 0.004, 5, 60)
print(f"Calibrated sigma: {sig_cal:.2f}")

ev["wp_base"] = wp(ev.margin, ev.t_left, 13.5)
ev["wp_cal"] = wp(ev.margin, ev.t_left, sig_cal)
ev.loc[ev.t_left <= 1e-6, ["wp_base", "wp_cal"]] = (ev.margin > 0).astype(float)

mins = ev[ev.t_left > 0.05]
for col in ["wp_base", "wp_cal"]:
    r = mins.loc[mins[col].idxmin()]
    print(f"{col}: min {r[col]*100:.3f}% (1 in {1/r[col]:,.0f}) at {r.elapsed_min:.2f} min, margin {r.margin}, {r.t_left:.2f} left")
for label, e in [("Half", 24), ("Low point", low.elapsed_min), ("End Q3", 36), ("9:33 Q4", 38 + 27 / 60)]:
    s = ev[ev.elapsed_min <= e + 1e-9].iloc[-1]
    print(f"{label:10s} margin {s.margin:4d}  base {wp(s.margin, 48 - e, 13.5)*100:.3f}%  cal {wp(s.margin, 48 - e, sig_cal)*100:.3f}%")
# last moments
print(ev.tail(6)[["elapsed_min", "team", "player", "pts", "sas", "nyk", "wp_base", "wp_cal"]].to_string())

# Output a series: start point plus one row per scoring event
start = pd.DataFrame([{"elapsed_min": 0, "sas": 0, "nyk": 0, "margin": 0, "wp_base": 0.5, "wp_cal": 0.5, "team": "", "player": "Tip-off", "pts": 0}])
out = pd.concat([start, ev[["elapsed_min", "sas", "nyk", "margin", "wp_base", "wp_cal", "team", "player", "pts"]]])
out.round(5).to_csv(sys.argv[3], index=False)
print("rows", len(out))
