# Fit the user's side-view screenshot (Blender Right Orthographic, nose to the left) to the model:
#   u = a * y + bu,  v = -a * z + bv   (pixels; model metres)
# from the hull's top and bottom silhouette, then convert the red strokes to model (y, z).
import json, sys
import numpy as np
from PIL import Image
S = r"C:\Users\PC\AppData\Local\Temp\claude\D--HADAL---The-Descent\1fbd6834-d7f7-4606-97c8-f13f3d7e4dae"
T = json.load(open(S + r"\scratchpad\shape_tables.json"))
tab = np.array(T["tab"])
Y, TOP, BOT = tab[:, 0], tab[:, 1], tab[:, 2]
im = np.asarray(Image.open(S + r"\images\5.webp").convert("RGB")).astype(int)
H, W, _ = im.shape
R, G, Bc = im[..., 0], im[..., 1], im[..., 2]
red = (R > 170) & (G < 90) & (Bc < 90)
lum = (R + G + Bc) / 3
hull = (lum > 92) & ~red
# silhouette per column: topmost and bottommost hull pixel (ignore the small header text, rows < 120 in the left part)
top_v, bot_v = {}, {}
for u in range(40, 1450, 4):
    col = hull[:, u].copy()
    col[:115] = False if u < 260 else col[:115]
    idx = np.nonzero(col)[0]
    if len(idx) < 20:
        continue
    top_v[u] = idx[0]
    bot_v[u] = idx[-1]

def resid(a, bu, bv, use_top=(40, 1150), use_bot=(300, 1450)):
    r = []
    for u, v in top_v.items():
        if use_top[0] <= u <= use_top[1]:
            y = (u - bu) / a
            if Y[0] <= y <= Y[-1]:
                r.append(v - (-a * np.interp(y, Y, TOP) + bv))
    for u, v in bot_v.items():
        if use_bot[0] <= u <= use_bot[1]:
            y = (u - bu) / a
            if Y[0] <= y <= Y[-1]:
                r.append(v - (-a * np.interp(y, Y, BOT) + bv))
    r = np.array(r)
    return r

def cost(p):
    r = resid(*p)
    return float(np.mean(np.minimum(r * r, 400.0)))      # robust: cap outliers (fittings, text)

p = np.array([417.0, 2060.0, 390.0])
best = (cost(p), p.copy())
step = np.array([20.0, 60.0, 30.0])
for it in range(200):
    improved = False
    for k in range(3):
        for s in (1, -1):
            q = best[1].copy(); q[k] += s * step[k]
            c = cost(q)
            if c < best[0]:
                best = (c, q); improved = True
    if not improved:
        step *= 0.5
        if step.max() < 0.01:
            break
a, bu, bv = best[1]
r = resid(a, bu, bv)
print("fit a=%.2f px/m  bu=%.1f  bv=%.1f  rms=%.2f px  n=%d  median|r|=%.2f" % (a, bu, bv, np.sqrt(best[0]), len(r), np.median(np.abs(r))))
# red strokes -> model coordinates
vs, us = np.nonzero(red)
pts = [((u - bu) / a, (bv - v) / a) for u, v in zip(us, vs)]
json.dump({"a": a, "bu": bu, "bv": bv, "red_yz": pts}, open(S + r"\scratchpad\side_fit.json", "w"))
# a check picture: the model's silhouette (blue) drawn from the fit over the screenshot, red kept
out = Image.open(S + r"\images\5.webp").convert("RGB")
px = out.load()
for y, t, b in zip(Y, TOP, BOT):
    u = int(round(a * y + bu))
    for z in (t, b):
        v = int(round(-a * z + bv))
        if 0 <= u < W and 0 <= v < H:
            for dv in (-1, 0, 1):
                if 0 <= v + dv < H:
                    px[u, v + dv] = (40, 160, 255)
out.save(S + r"\scratchpad\side_fit_check.png")
print("ok")