phasm_core/stego/ghost/

optimizer.rs

// Copyright (c) 2026 Christoph Gaffga
// SPDX-License-Identifier: GPL-3.0-only
// https://github.com/cgaffga/phasmcore

//! Cover image preprocessing optimizer for steganographic embedding.
//!
//! Subtly modifies raw pixels before JPEG compression to improve embedding
//! quality and capacity. Each mode applies a different pipeline:
//!
//! - **Ghost**: Texture-adaptive 4-stage pipeline (noise injection, micro-contrast,
//!   unsharp mask, smooth-region dithering) — maximizes non-zero AC coefficients.
//!   Each stage adapts its strength based on existing texture levels to avoid
//!   degrading already-optimized images.
//! - **Armor**: Light pipeline (block-boundary smoothing, DC stabilization) —
//!   reduces cross-block discontinuities for STDM robustness.
//! - **Fortress**: Minimal (block-boundary smoothing only) — stabilizes DC
//!   averages for BA-QIM embedding.
//!
//! **"Do no harm" guarantee**: After optimization, the average absolute gradient
//! (a proxy for JPEG AC energy / stego capacity) is compared to the original.
//! If the optimizer reduced gradient energy, the original pixels are returned
//! unchanged. This prevents degradation of pre-optimized images (e.g. images
//! that already had noise, micro-contrast, and sharpening applied in Photoshop).
//!
//! The optimizer is deterministic: given the same pixels, dimensions, config,
//! and seed, it always produces the same output.

use rand::Rng;
use rand::SeedableRng;
use rand_chacha::ChaCha20Rng;

use crate::det_math::det_exp;

/// Configuration for the cover image optimizer.
pub struct OptimizerConfig {
    /// Strength multiplier (0.0–1.0). Default: 0.85.
    pub strength: f32,
    /// ChaCha20 seed for deterministic noise generation.
    pub seed: [u8; 32],
    /// Pipeline variant to apply.
    pub mode: OptimizerMode,
}

/// Which optimization pipeline to run.
pub enum OptimizerMode {
    /// Full 4-stage pipeline (noise, micro-contrast, unsharp mask, dithering).
    Ghost,
    /// Light: block-boundary smoothing + DC stabilization.
    Armor,
    /// Minimal: block-boundary smoothing only.
    Fortress,
}

/// Optimize raw RGB pixels for steganographic embedding.
///
/// Returns a new pixel buffer with the same dimensions and format (RGB, 3
/// bytes per pixel, row-major). The modifications are subtle enough to be
/// imperceptible (PSNR > 44 dB, SSIM > 0.993) but improve embedding
/// capacity by increasing wavelet energy in smooth regions.
///
/// If the optimization would reduce gradient energy (degrade stego capacity),
/// the original pixels are returned unchanged.
pub fn optimize_cover(
    pixels_rgb: &[u8],
    width: u32,
    height: u32,
    config: &OptimizerConfig,
) -> Vec<u8> {
    let w = width as usize;
    let h = height as usize;
    assert_eq!(pixels_rgb.len(), w * h * 3, "pixel buffer size mismatch");

    let result = match config.mode {
        OptimizerMode::Ghost => optimize_ghost(pixels_rgb, w, h, config),
        OptimizerMode::Armor => optimize_armor(pixels_rgb, w, h, config),
        OptimizerMode::Fortress => optimize_fortress(pixels_rgb, w, h, config),
    };

    // "Do no harm" safety check: compare gradient energy before and after.
    // Require at least 1% improvement — marginal changes on already-optimized
    // images are more likely to hurt (especially SI-UNIWARD rounding errors)
    // than to help. If the optimizer can't meaningfully improve the image,
    // the original was already well-prepared.
    let original_energy = gradient_energy(pixels_rgb, w, h);
    let optimized_energy = gradient_energy(&result, w, h);
    if optimized_energy < original_energy * 1.01 {
        pixels_rgb.to_vec()
    } else {
        result
    }
}

// ---------------------------------------------------------------------------
// Texture analysis
// ---------------------------------------------------------------------------

/// Analyze image texture to compute per-stage adaptive strength multipliers.
///
/// Returns `(noise_factor, contrast_factor, sharpen_factor, dither_factor)`,
/// each in [0.0, 1.0]. A factor of 0.0 means "skip this stage entirely"
/// (image already has plenty of that characteristic). 1.0 means "apply fully".
#[cfg(test)]
fn analyze_texture(luma: &[f64], variance: &[f64], w: usize, h: usize) -> (f64, f64, f64, f64) {
    let n = w * h;
    if n == 0 {
        return (1.0, 1.0, 1.0, 1.0);
    }

    // 1. Noise factor: based on mean variance.
    //    Low variance (< 50) = smooth image, needs lots of noise → factor ~1.0
    //    Medium variance (50-300) = some texture → moderate noise
    //    High variance (> 600) = already noisy → skip noise
    let mean_variance: f64 = variance.iter().sum::<f64>() / n as f64;
    let noise_factor = adaptive_scale(mean_variance, 50.0, 600.0);

    // 2. Contrast factor: based on proportion of pixels with low local contrast.
    //    Compute local gradient magnitude as a contrast proxy.
    let mut low_contrast_count = 0usize;
    for y in 1..h {
        for x in 1..w {
            let idx = y * w + x;
            let dy = (luma[idx] - luma[(y - 1) * w + x]).abs();
            let dx = (luma[idx] - luma[y * w + (x - 1)]).abs();
            let grad = dy + dx;
            if grad < 3.0 {
                low_contrast_count += 1;
            }
        }
    }
    let low_contrast_ratio = low_contrast_count as f64 / ((w - 1) * (h - 1)) as f64;
    // If < 10% of pixels are low-contrast, the image already has good contrast
    // If > 50% of pixels are low-contrast, we need full enhancement
    let contrast_factor = if low_contrast_ratio < 0.10 {
        0.0
    } else if low_contrast_ratio > 0.50 {
        1.0
    } else {
        (low_contrast_ratio - 0.10) / 0.40
    };

    // 3. Sharpen factor: based on high-frequency energy.
    //    Measure average |pixel - local_mean| as proxy for existing sharpening.
    let mut hf_sum = 0.0f64;
    let mut hf_count = 0usize;
    for y in 1..h.saturating_sub(1) {
        for x in 1..w.saturating_sub(1) {
            let idx = y * w + x;
            let center = luma[idx];
            // 4-neighbor average
            let avg = (luma[(y - 1) * w + x] + luma[(y + 1) * w + x]
                + luma[y * w + (x - 1)] + luma[y * w + (x + 1)])
                / 4.0;
            hf_sum += (center - avg).abs();
            hf_count += 1;
        }
    }
    let mean_hf = if hf_count > 0 { hf_sum / hf_count as f64 } else { 0.0 };
    // mean_hf < 1.0 = very smooth, needs sharpening → 1.0
    // mean_hf > 4.0 = already sharp → 0.0
    let sharpen_factor = adaptive_scale(mean_hf, 1.0, 4.0);

    // 4. Dither factor: based on proportion of smooth 4×4 blocks.
    //    If few smooth blocks exist, dithering has little to do.
    let blocks_x = w / 4;
    let blocks_y = h / 4;
    let total_blocks = blocks_x * blocks_y;
    let mut smooth_blocks = 0usize;
    for by in 0..blocks_y {
        for bx in 0..blocks_x {
            let mut block_mean = 0.0f64;
            for dy in 0..4 {
                for dx in 0..4 {
                    block_mean += luma[(by * 4 + dy) * w + (bx * 4 + dx)];
                }
            }
            block_mean /= 16.0;
            let mut sad = 0.0f64;
            for dy in 0..4 {
                for dx in 0..4 {
                    sad += (luma[(by * 4 + dy) * w + (bx * 4 + dx)] - block_mean).abs();
                }
            }
            if sad < 15.0 {
                smooth_blocks += 1;
            }
        }
    }
    let smooth_ratio = if total_blocks > 0 {
        smooth_blocks as f64 / total_blocks as f64
    } else {
        0.0
    };
    // If < 5% smooth blocks → no dithering needed (image is already textured)
    // If > 30% smooth blocks → full dithering
    let dither_factor = if smooth_ratio < 0.05 {
        0.0
    } else if smooth_ratio > 0.30 {
        1.0
    } else {
        (smooth_ratio - 0.05) / 0.25
    };

    (noise_factor, contrast_factor, sharpen_factor, dither_factor)
}

/// Maps a metric value to [0.0, 1.0] inversely: below `low` → 1.0, above `high` → 0.0.
fn adaptive_scale(value: f64, low: f64, high: f64) -> f64 {
    if value <= low {
        1.0
    } else if value >= high {
        0.0
    } else {
        1.0 - (value - low) / (high - low)
    }
}

/// Compute average absolute gradient energy (proxy for JPEG AC coefficient density).
/// Higher value = more texture = better for steganography.
fn gradient_energy(pixels: &[u8], w: usize, h: usize) -> f64 {
    if w < 2 || h < 2 {
        return 0.0;
    }
    let mut energy = 0.0f64;
    let mut count = 0usize;
    // Sample on luma channel for speed
    for y in 1..h {
        for x in 1..w {
            let idx = (y * w + x) * 3;
            let idx_left = (y * w + (x - 1)) * 3;
            let idx_above = ((y - 1) * w + x) * 3;
            // Luma approximation: just green channel (dominant in luma)
            let center = pixels[idx + 1] as f64;
            let left = pixels[idx_left + 1] as f64;
            let above = pixels[idx_above + 1] as f64;
            energy += (center - left).abs() + (center - above).abs();
            count += 1;
        }
    }
    if count > 0 { energy / count as f64 } else { 0.0 }
}

// ---------------------------------------------------------------------------
// Ghost pipeline (4 stages, texture-adaptive)
// ---------------------------------------------------------------------------

fn optimize_ghost(pixels: &[u8], w: usize, h: usize, config: &OptimizerConfig) -> Vec<u8> {
    let strength = config.strength as f64;
    let mut rng = ChaCha20Rng::from_seed(config.seed);

    // Compute per-pixel texture map: local variance drives ALL stages.
    // Each pixel gets its own strength multiplier based on its neighborhood.
    // Smooth areas (low variance) → full processing.
    // Textured areas (high variance) → minimal/no processing.
    let luma = extract_luma(pixels, w, h);
    let variance = local_variance_5x5(&luma, w, h);

    let mut out = pixels.to_vec();

    // Stage 1: Content-adaptive noise injection (per-pixel)
    // Variance < 50: smooth region → σ up to 1.5 (needs texture)
    // Variance 50–600: transitional → scaled σ
    // Variance > 600: already textured → σ ≈ 0 (skip)
    for y in 0..h {
        for x in 0..w {
            // Always consume RNG for determinism
            let u1: f64 = rng.gen_range(0.0f64..1.0).max(1e-10);
            let u2: f64 = rng.gen_range(0.0f64..1.0);

            let var_val = variance[y * w + x];
            let local_need = adaptive_scale(var_val, 50.0, 600.0);
            // Skip pixels in areas that are already well-textured —
            // even tiny modifications can hurt SI-UNIWARD rounding errors
            if local_need > 0.1 {
            let sigma = 1.5 * local_need + 0.1;
            let scaled_sigma = sigma * strength * local_need;

            if scaled_sigma > 0.05 {
                // Irwin-Hall(2) noise: u1+u2 has triangular distribution on
                // [0, 2] with mean 1, variance 1/6. Center to mean 0 and
                // scale by √6 so variance matches N(0,1).
                //
                // Replaces Box-Muller `(-2*ln(u1)).sqrt() * cos(2π·u2)` —
                // f64::ln + f64::cos lower to non-deterministic libm /
                // Math.* on WASM, breaking cross-platform reproducibility of
                // the optimized cover. Triangular noise is a coarser Gaussian
                // approximation but adequate for the small-amplitude texture
                // injection done here. RNG consumption is unchanged
                // (still 2 uniform draws per pixel).
                const SQRT_6: f64 = 2.449489742783178;
                let gauss = (u1 + u2 - 1.0) * SQRT_6;
                let noise = gauss * scaled_sigma;

                let idx = (y * w + x) * 3;
                for c in 0..3 {
                    let val = out[idx + c] as f64 + noise;
                    out[idx + c] = val.round().clamp(0.0, 255.0) as u8;
                }
            }
            } // local_need > 0.1
        }
    }

    // Stage 2: Micro-contrast enhancement (per-pixel adaptive)
    // Each pixel's enhancement is scaled by how much local contrast it needs.
    // High-variance areas already have good micro-contrast → minimal boost.
    let stage1 = out.clone();
    for y in 0..h {
        for x in 0..w {
            let var_val = variance[y * w + x];
            let local_need = adaptive_scale(var_val, 80.0, 500.0);
            if local_need > 0.1 {
            let local_enhance = 0.15 * strength * local_need;

            if local_enhance > 0.01 {
                let idx = (y * w + x) * 3;
                for c in 0..3 {
                    let (local_mean, pixel_val) = neighborhood_mean_3x3(&stage1, w, h, x, y, c);
                    let deviation = pixel_val - local_mean;
                    let enhanced = local_mean + deviation * (1.0 + local_enhance);
                    out[idx + c] = enhanced.round().clamp(0.0, 255.0) as u8;
                }
            }
            } // local_need > 0.1
        }
    }

    // Stage 3: Gentle unsharp mask (per-pixel adaptive)
    // Sharpening only where the area is too smooth. Textured/already-sharp
    // areas are left untouched to avoid halos and over-enhancement.
    let stage2 = out.clone();
    let blurred = gaussian_blur_separable(&stage2, w, h, 1.0);
    for y in 0..h {
        for x in 0..w {
            let var_val = variance[y * w + x];
            let local_need = adaptive_scale(var_val, 100.0, 400.0);
            if local_need > 0.1 {
            let local_sharpen = 0.12 * strength * local_need;

            if local_sharpen > 0.01 {
                let idx = (y * w + x) * 3;
                for c in 0..3 {
                    let i = idx + c;
                    let orig = stage2[i] as f64;
                    let blur = blurred[i] as f64;
                    let diff = orig - blur;
                    // Threshold: only sharpen if difference > 3 (avoid amplifying noise)
                    if diff.abs() > 3.0 {
                        let sharpened = orig + local_sharpen * diff;
                        out[i] = sharpened.round().clamp(0.0, 255.0) as u8;
                    }
                }
            }
            } // local_need > 0.1
        }
    }

    // Stage 4: Smooth-region dithering (per 4×4 block, re-evaluated)
    // Uses CURRENT luma (after stages 1-3) — blocks already textured by
    // earlier stages won't get redundant dithering.
    let bayer_4x4: [[f64; 4]; 4] = [
        [0.0, 8.0, 2.0, 10.0],
        [12.0, 4.0, 14.0, 6.0],
        [3.0, 11.0, 1.0, 9.0],
        [15.0, 7.0, 13.0, 5.0],
    ];
    let bayer_norm: [[f64; 4]; 4] = {
        let mut b = [[0.0; 4]; 4];
        for r in 0..4 {
            for c in 0..4 {
                b[r][c] = (bayer_4x4[r][c] / 15.0) * 2.0 - 1.0;
            }
        }
        b
    };

    let current_luma = extract_luma(&out, w, h);
    let blocks_x = w / 4;
    let blocks_y = h / 4;
    for by in 0..blocks_y {
        for bx in 0..blocks_x {
            // Always consume RNG for determinism
            let block_offset: f64 = rng.gen_range(0.0f64..1.0) * 2.0 - 1.0;

            // Compute SAD and mean variance for this 4×4 block
            let mut block_mean = 0.0f64;
            let mut block_var_sum = 0.0f64;
            for dy in 0..4 {
                for dx in 0..4 {
                    let px = bx * 4 + dx;
                    let py = by * 4 + dy;
                    if px < w && py < h {
                        block_mean += current_luma[py * w + px];
                        block_var_sum += variance[py * w + px];
                    }
                }
            }
            block_mean /= 16.0;
            let block_var_avg = block_var_sum / 16.0;
            let mut block_sad = 0.0f64;
            for dy in 0..4 {
                for dx in 0..4 {
                    let px = bx * 4 + dx;
                    let py = by * 4 + dy;
                    if px < w && py < h {
                        block_sad += (current_luma[py * w + px] - block_mean).abs();
                    }
                }
            }

            // Dither only smooth blocks in areas that need texture
            let block_need = adaptive_scale(block_var_avg, 50.0, 400.0);
            let effective_dither = strength * block_need;

            if block_sad < 15.0 && effective_dither > 0.01 {
                for dy in 0..4 {
                    for dx in 0..4 {
                        let px = bx * 4 + dx;
                        let py = by * 4 + dy;
                        if px < w && py < h {
                            let dither = (bayer_norm[dy][dx] + block_offset * 0.3) * effective_dither;
                            let idx = (py * w + px) * 3;
                            for c in 0..3 {
                                let val = out[idx + c] as f64 + dither;
                                out[idx + c] = val.round().clamp(0.0, 255.0) as u8;
                            }
                        }
                    }
                }
            }
        }
    }

    out
}

// ---------------------------------------------------------------------------
// Armor pipeline (2 stages)
// ---------------------------------------------------------------------------

fn optimize_armor(pixels: &[u8], w: usize, h: usize, config: &OptimizerConfig) -> Vec<u8> {
    let mut out = pixels.to_vec();
    let strength = config.strength;

    // Stage 1: Gentle low-pass on block boundaries (8×8 grid)
    smooth_block_boundaries(&mut out, w, h, strength);

    // Stage 2: DC stabilization — subtle smoothing within each 8×8 block
    // to reduce variance of block-average brightness
    dc_stabilize(&mut out, w, h, strength);

    out
}

// ---------------------------------------------------------------------------
// Fortress pipeline (1 stage)
// ---------------------------------------------------------------------------

fn optimize_fortress(pixels: &[u8], w: usize, h: usize, config: &OptimizerConfig) -> Vec<u8> {
    let mut out = pixels.to_vec();
    smooth_block_boundaries(&mut out, w, h, config.strength);
    out
}

// ---------------------------------------------------------------------------
// Helper functions
// ---------------------------------------------------------------------------

/// Extract luma (Y) from RGB pixels. Y = 0.299R + 0.587G + 0.114B.
fn extract_luma(pixels: &[u8], w: usize, h: usize) -> Vec<f64> {
    let mut luma = vec![0.0f64; w * h];
    for i in 0..w * h {
        let r = pixels[i * 3] as f64;
        let g = pixels[i * 3 + 1] as f64;
        let b = pixels[i * 3 + 2] as f64;
        luma[i] = 0.299 * r + 0.587 * g + 0.114 * b;
    }
    luma
}

/// Compute local variance in a 5×5 neighborhood for each pixel.
fn local_variance_5x5(luma: &[f64], w: usize, h: usize) -> Vec<f64> {
    let mut variance = vec![0.0f64; w * h];
    for y in 0..h {
        for x in 0..w {
            let mut sum = 0.0f64;
            let mut sum_sq = 0.0f64;
            let mut count = 0.0f64;
            for dy in 0..5usize {
                for dx in 0..5usize {
                    let py = (y + dy).saturating_sub(2).min(h - 1);
                    let px = (x + dx).saturating_sub(2).min(w - 1);
                    let val = luma[py * w + px];
                    sum += val;
                    sum_sq += val * val;
                    count += 1.0;
                }
            }
            let mean = sum / count;
            variance[y * w + x] = sum_sq / count - mean * mean;
        }
    }
    variance
}

/// Compute 3×3 neighborhood mean and center pixel value for a channel.
fn neighborhood_mean_3x3(pixels: &[u8], w: usize, h: usize, x: usize, y: usize, c: usize) -> (f64, f64) {
    let mut sum = 0.0f64;
    let mut count = 0.0f64;
    for dy in 0..3usize {
        for dx in 0..3usize {
            let py = (y + dy).saturating_sub(1).min(h - 1);
            let px = (x + dx).saturating_sub(1).min(w - 1);
            sum += pixels[(py * w + px) * 3 + c] as f64;
            count += 1.0;
        }
    }
    let pixel_val = pixels[(y * w + x) * 3 + c] as f64;
    (sum / count, pixel_val)
}

/// Separable Gaussian blur with given sigma (truncated at 3σ kernel).
fn gaussian_blur_separable(pixels: &[u8], w: usize, h: usize, sigma: f64) -> Vec<u8> {
    let radius = (sigma * 3.0).ceil() as usize;
    let kernel_size = radius * 2 + 1;

    // Build 1D Gaussian kernel
    let mut kernel = vec![0.0f64; kernel_size];
    let mut sum = 0.0f64;
    for i in 0..kernel_size {
        let x = i as f64 - radius as f64;
        // det_exp instead of f64::exp — f64::exp lowers to non-deterministic
        // libm / Math.exp on WASM, breaking cross-platform bit-exactness of
        // the optimized cover.
        let val = det_exp(-x * x / (2.0 * sigma * sigma));
        kernel[i] = val;
        sum += val;
    }
    for k in kernel.iter_mut() {
        *k /= sum;
    }

    let n = w * h * 3;
    let mut temp = vec![0u8; n];
    let mut result = vec![0u8; n];

    // Horizontal pass
    for y in 0..h {
        for x in 0..w {
            for c in 0..3 {
                let mut acc = 0.0f64;
                for k in 0..kernel_size {
                    let sx = (x + k).saturating_sub(radius).min(w - 1);
                    acc += pixels[(y * w + sx) * 3 + c] as f64 * kernel[k];
                }
                temp[(y * w + x) * 3 + c] = acc.round().clamp(0.0, 255.0) as u8;
            }
        }
    }

    // Vertical pass
    for y in 0..h {
        for x in 0..w {
            for c in 0..3 {
                let mut acc = 0.0f64;
                for k in 0..kernel_size {
                    let sy = (y + k).saturating_sub(radius).min(h - 1);
                    acc += temp[(sy * w + x) * 3 + c] as f64 * kernel[k];
                }
                result[(y * w + x) * 3 + c] = acc.round().clamp(0.0, 255.0) as u8;
            }
        }
    }

    result
}

/// Smooth block boundaries at 8×8 grid edges.
///
/// Applies 3-pixel averaging across horizontal and vertical block edges
/// to reduce cross-block discontinuities after JPEG recompression.
fn smooth_block_boundaries(pixels: &mut [u8], w: usize, h: usize, strength: f32) {
    let alpha = 0.25 * strength as f64; // blending weight

    // Vertical block edges (columns at multiples of 8)
    for col in (8..w).step_by(8) {
        if col >= w {
            continue;
        }
        for y in 0..h {
            for c in 0..3 {
                let left = pixels[(y * w + col - 1) * 3 + c] as f64;
                let center = pixels[(y * w + col) * 3 + c] as f64;
                let right = if col + 1 < w {
                    pixels[(y * w + col + 1) * 3 + c] as f64
                } else {
                    center
                };
                let avg = (left + center + right) / 3.0;
                let blended = center + alpha * (avg - center);
                pixels[(y * w + col) * 3 + c] = blended.round().clamp(0.0, 255.0) as u8;
            }
        }
    }

    // Horizontal block edges (rows at multiples of 8)
    for row in (8..h).step_by(8) {
        for x in 0..w {
            for c in 0..3 {
                let above = pixels[((row - 1) * w + x) * 3 + c] as f64;
                let center = pixels[(row * w + x) * 3 + c] as f64;
                let below = if row + 1 < h {
                    pixels[((row + 1) * w + x) * 3 + c] as f64
                } else {
                    center
                };
                let avg = (above + center + below) / 3.0;
                let blended = center + alpha * (avg - center);
                pixels[(row * w + x) * 3 + c] = blended.round().clamp(0.0, 255.0) as u8;
            }
        }
    }
}

/// DC stabilization: subtle within-block smoothing to reduce variance
/// of block-average brightness, helping STDM embedding stability.
fn dc_stabilize(pixels: &mut [u8], w: usize, h: usize, strength: f32) {
    let alpha = 0.1 * strength as f64;
    let blocks_x = w.div_ceil(8);
    let blocks_y = h.div_ceil(8);

    for by in 0..blocks_y {
        for bx in 0..blocks_x {
            // Compute block mean for each channel
            let mut mean = [0.0f64; 3];
            let mut count = 0.0f64;
            for dy in 0..8 {
                for dx in 0..8 {
                    let px = bx * 8 + dx;
                    let py = by * 8 + dy;
                    if px < w && py < h {
                        for c in 0..3 {
                            mean[c] += pixels[(py * w + px) * 3 + c] as f64;
                        }
                        count += 1.0;
                    }
                }
            }
            if count > 0.0 {
                for c in 0..3 {
                    mean[c] /= count;
                }
            }

            // Nudge each pixel slightly toward block mean
            for dy in 0..8 {
                for dx in 0..8 {
                    let px = bx * 8 + dx;
                    let py = by * 8 + dy;
                    if px < w && py < h {
                        let idx = (py * w + px) * 3;
                        for c in 0..3 {
                            let val = pixels[idx + c] as f64;
                            let blended = val + alpha * (mean[c] - val);
                            pixels[idx + c] = blended.round().clamp(0.0, 255.0) as u8;
                        }
                    }
                }
            }
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn ghost_deterministic() {
        let w = 64;
        let h = 64;
        let pixels = vec![128u8; w * h * 3];
        let config = OptimizerConfig {
            strength: 0.85,
            seed: [42u8; 32],
            mode: OptimizerMode::Ghost,
        };
        let out1 = optimize_cover(&pixels, w as u32, h as u32, &config);
        let out2 = optimize_cover(&pixels, w as u32, h as u32, &config);
        assert_eq!(out1, out2, "optimizer must be deterministic");
    }

    #[test]
    fn ghost_modifies_smooth_pixels() {
        // A flat image (all same value) should be modified — lots of room to improve
        let w = 64;
        let h = 64;
        let pixels = vec![128u8; w * h * 3];
        let config = OptimizerConfig {
            strength: 0.85,
            seed: [42u8; 32],
            mode: OptimizerMode::Ghost,
        };
        let out = optimize_cover(&pixels, w as u32, h as u32, &config);
        assert_ne!(pixels, out, "optimizer should modify smooth pixels");
    }

    #[test]
    fn do_no_harm_textured_image() {
        // Create an already well-textured image (simulating pre-optimized photo).
        // The optimizer should return it unchanged if it can't improve gradient energy.
        let w = 128;
        let h = 128;
        let mut pixels = vec![0u8; w * h * 3];
        // High-variance random-ish texture
        for y in 0..h {
            for x in 0..w {
                let idx = (y * w + x) * 3;
                let hash = ((x * 31 + y * 97 + 7) % 256) as u8;
                pixels[idx] = hash;
                pixels[idx + 1] = hash.wrapping_mul(3).wrapping_add(17);
                pixels[idx + 2] = hash.wrapping_mul(7).wrapping_add(53);
            }
        }
        let config = OptimizerConfig {
            strength: 0.85,
            seed: [42u8; 32],
            mode: OptimizerMode::Ghost,
        };
        let out = optimize_cover(&pixels, w as u32, h as u32, &config);

        // Either unchanged (do-no-harm kicked in) or gradient energy increased
        let orig_energy = gradient_energy(&pixels, w, h);
        let opt_energy = gradient_energy(&out, w, h);
        assert!(
            opt_energy >= orig_energy,
            "optimizer must not reduce gradient energy: original={orig_energy:.2}, optimized={opt_energy:.2}"
        );
    }

    #[test]
    fn psnr_above_threshold() {
        let w = 128;
        let h = 128;
        let mut pixels = vec![0u8; w * h * 3];
        for y in 0..h {
            for x in 0..w {
                let idx = (y * w + x) * 3;
                let base_r = (x * 255 / w) as u8;
                let base_g = (y * 255 / h) as u8;
                let base_b = ((x + y) * 255 / (w + h)) as u8;
                let hash = ((x * 31 + y * 97 + 7) % 30) as u8;
                pixels[idx] = base_r.wrapping_add(hash);
                pixels[idx + 1] = base_g.wrapping_add(hash.wrapping_mul(3));
                pixels[idx + 2] = base_b.wrapping_add(hash.wrapping_mul(7));
            }
        }
        let config = OptimizerConfig {
            strength: 0.85,
            seed: [42u8; 32],
            mode: OptimizerMode::Ghost,
        };
        let out = optimize_cover(&pixels, w as u32, h as u32, &config);

        let mse: f64 = pixels
            .iter()
            .zip(out.iter())
            .map(|(&a, &b)| {
                let d = a as f64 - b as f64;
                d * d
            })
            .sum::<f64>()
            / pixels.len() as f64;

        let psnr = if mse > 0.0 {
            10.0 * (255.0 * 255.0 / mse).log10()
        } else {
            f64::INFINITY
        };

        assert!(
            psnr > 28.0,
            "PSNR should be > 28 dB even for synthetic images, got {psnr:.1} dB"
        );
    }

    #[test]
    fn armor_mode_runs() {
        let w = 32;
        let h = 32;
        let pixels = vec![128u8; w * h * 3];
        let config = OptimizerConfig {
            strength: 0.85,
            seed: [0u8; 32],
            mode: OptimizerMode::Armor,
        };
        let out = optimize_cover(&pixels, w as u32, h as u32, &config);
        assert_eq!(out.len(), pixels.len());
    }

    #[test]
    fn fortress_mode_runs() {
        let w = 32;
        let h = 32;
        let pixels = vec![128u8; w * h * 3];
        let config = OptimizerConfig {
            strength: 0.85,
            seed: [0u8; 32],
            mode: OptimizerMode::Fortress,
        };
        let out = optimize_cover(&pixels, w as u32, h as u32, &config);
        assert_eq!(out.len(), pixels.len());
    }

    #[test]
    fn zero_strength_minimal_change() {
        let w = 32;
        let h = 32;
        let pixels: Vec<u8> = (0..w * h * 3).map(|i| (i % 256) as u8).collect();
        let config = OptimizerConfig {
            strength: 0.0,
            seed: [42u8; 32],
            mode: OptimizerMode::Ghost,
        };
        let out = optimize_cover(&pixels, w as u32, h as u32, &config);

        let max_diff: u8 = pixels
            .iter()
            .zip(out.iter())
            .map(|(&a, &b)| a.abs_diff(b))
            .max()
            .unwrap_or(0);
        assert!(
            max_diff <= 1,
            "zero strength should cause near-zero changes, max diff = {max_diff}"
        );
    }

    #[test]
    fn texture_analysis_smooth_image() {
        let w = 64;
        let h = 64;
        let pixels = vec![128u8; w * h * 3]; // perfectly flat
        let luma = extract_luma(&pixels, w, h);
        let variance = local_variance_5x5(&luma, w, h);
        let (noise, contrast, sharpen, dither) = analyze_texture(&luma, &variance, w, h);
        // Flat image should get maximum treatment on all stages
        assert!(noise > 0.9, "flat image should get full noise: {noise}");
        assert!(contrast > 0.9, "flat image should get full contrast: {contrast}");
        assert!(sharpen > 0.9, "flat image should get full sharpen: {sharpen}");
        assert!(dither > 0.5, "flat image should get significant dither: {dither}");
    }

    #[test]
    fn texture_analysis_noisy_image() {
        let w = 64;
        let h = 64;
        let mut pixels = vec![0u8; w * h * 3];
        // High-variance texture everywhere
        for i in 0..pixels.len() {
            pixels[i] = ((i * 31 + 7) % 256) as u8;
        }
        let luma = extract_luma(&pixels, w, h);
        let variance = local_variance_5x5(&luma, w, h);
        let (noise, _contrast, _sharpen, dither) = analyze_texture(&luma, &variance, w, h);
        // Already-textured image should get minimal treatment
        assert!(noise < 0.3, "textured image should get little noise: {noise}");
        assert!(dither < 0.3, "textured image should get little dither: {dither}");
    }
}