"""Offline unit tests: no network, no models. Exercise the pure-Python core."""

from __future__ import annotations

import numpy as np
from PIL import Image

from veil_pgd.fitness.embed import cosine
from veil_pgd.fitness.objective import combined_fitness, stealth_penalty
from veil_pgd.render import render
from veil_pgd.stealth import evaluate_stealth
from veil_pgd.stealth.gate import StealthThresholds
from veil_pgd.types import RenderSpec, StealthReport
from veil_pgd.util.labels import normalize_label, parse_label_json


def _img(seed=0):
    rng = np.random.default_rng(seed)
    return Image.fromarray(rng.integers(90, 140, (256, 256, 3)).astype("uint8"))


def test_normalize_label():
    assert normalize_label("  The Golden Retriever! ") == "golden retriever"


def test_parse_label_json():
    assert parse_label_json('{"label": "Fire Hydrant"}') == "fire hydrant"
    assert parse_label_json('noise {"label":"cat"} more') == "cat"
    assert parse_label_json("no json here") is None


def test_cosine():
    assert abs(cosine([1, 0], [1, 0]) - 1.0) < 1e-9
    assert abs(cosine([1, 0], [0, 1])) < 1e-9


def test_render_and_stealth_faint_passes_gate():
    img = _img()
    spec = RenderSpec(text="fire hydrant", font_px=14, alpha=0.2)
    out = render(img, spec)
    assert out.size == img.size
    rep = evaluate_stealth(img, out)  # no lpips_fn -> lpips skipped
    assert rep.psnr > 30
    assert rep.passed


def test_high_opacity_hurts_stealth():
    img = _img()
    faint = evaluate_stealth(img, render(img, RenderSpec(text="x" * 20, font_px=14, alpha=0.15)))
    loud = evaluate_stealth(img, render(img, RenderSpec(text="x" * 20, font_px=24, alpha=1.0,
                                                        color_strategy="fixed",
                                                        fixed_rgb=(255, 0, 0))))
    assert loud.psnr < faint.psnr


def test_combined_fitness_penalizes_gate_failures():
    th = StealthThresholds()
    good = StealthReport(psnr=40, ssim=0.99, lpips=0.01, delta_e_p95=1.0, passed=True)
    bad = StealthReport(psnr=20, ssim=0.5, lpips=0.5, delta_e_p95=10.0, passed=False)
    assert stealth_penalty(good, th) == 0.0
    assert stealth_penalty(bad, th) > 0.0
    f_good = combined_fitness(0.8, 0.9, good, th)
    f_bad = combined_fitness(0.8, 0.9, bad, th)
    assert f_good > f_bad