stet-graphics 0.2.1

// stet - A PostScript Interpreter
// Copyright (c) 2026 Scott Bowman
// SPDX-License-Identifier: Apache-2.0 OR MIT

//! ICC color profile support via moxcms.
//!
//! Parses embedded ICC profiles from `[/ICCBased stream]` color spaces and
//! converts colors to sRGB. Also searches for system CMYK profiles to improve
//! DeviceCMYK → RGB conversion beyond the naive PLRM formula.

pub mod bpc;
mod perceptual;

use bpc::{
    BpcParams, apply_bpc_f64, apply_bpc_rgb_u8, compute_bpc_params, detect_source_black_point,
};
use moxcms::{
    CmsError, ColorProfile, DataColorSpace, Layout, RenderingIntent, TransformExecutor,
    TransformOptions,
};
use std::collections::HashMap;
use std::sync::Arc;

/// Re-export so callers in `stet-pdf-reader` (and other consumers
/// without a direct moxcms dependency) can specify the rendering intent
/// for [`IccCache::convert_color_with_intent`] etc. without taking a
/// moxcms dep themselves.
pub use moxcms::RenderingIntent as IccRenderingIntent;

/// Map a PDF gstate `rendering_intent` byte (0=Perceptual, 1=RelCol,
/// 2=Saturation, 3=AbsCol — the encoding used by
/// `stet-pdf-reader::content::graphics_state::PdfGraphicsState`) to the
/// corresponding [`IccRenderingIntent`]. Unknown values fall back to
/// Perceptual, matching the PDF spec's default.
#[inline]
pub fn intent_from_pdf_byte(b: u8) -> IccRenderingIntent {
    match b {
        1 => IccRenderingIntent::RelativeColorimetric,
        2 => IccRenderingIntent::Saturation,
        3 => IccRenderingIntent::AbsoluteColorimetric,
        _ => IccRenderingIntent::Perceptual,
    }
}

/// SHA-256 hash used as profile key.
pub type ProfileHash = [u8; 32];

/// Cache key for per-intent single-colour lookups: `(hash-prefix,
/// quantized_components, intent_byte)`.
type IntentColorKey = (u64, [u16; 4], u8);

/// Black Point Compensation mode for CMYK→sRGB conversion.
///
/// Reference renderers (Ghostscript, Acrobat, Firefox via lcms2) apply BPC by
/// default for relative-colorimetric CMYK→sRGB. moxcms 0.8.1 ships BPC
/// commented out, so without it K-heavy colors render visibly lighter than
/// reference renderers. See `docs/PLAN-BPC.md` for the full design.
#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
pub enum BpcMode {
    /// Skip BPC; matches stet's pre-fix behavior. Useful for proofing-style
    /// renders that should preserve actual densities, or for bit-for-bit
    /// reproduction of older baselines.
    Off,
    /// Always apply BPC during CMYK→sRGB conversion.
    On,
    /// Default — currently equivalent to `On`. Reserved for forward
    /// compatibility (eventually could honor PDF rendering-intent or
    /// output-intent hints).
    #[default]
    Auto,
}

impl BpcMode {
    /// True when BPC should be applied at conversion time.
    #[inline]
    pub fn is_enabled(self) -> bool {
        matches!(self, BpcMode::On | BpcMode::Auto)
    }
}

/// Construction-time options for [`IccCache`].
///
/// Bundles together the BPC mode and an optional pre-supplied source CMYK
/// profile (overriding the automatic system-profile search). Created via
/// [`IccCache::new_with_options`].
#[derive(Clone, Default)]
pub struct IccCacheOptions {
    /// BPC mode for CMYK→sRGB conversion.
    pub bpc_mode: BpcMode,
    /// Raw bytes of a source CMYK profile to register as the system profile.
    /// When `None`, the cache is created empty and the caller is responsible
    /// for invoking [`IccCache::search_system_cmyk_profile`] (or providing
    /// bytes some other way).
    pub source_cmyk_profile: Option<Vec<u8>>,
}

/// Identity Gray→RGB transform: maps each gray value to equal R=G=B.
/// Used as fallback when a Gray ICC profile can't produce a proper transform.
struct GrayToRgbIdentity;

impl TransformExecutor<u8> for GrayToRgbIdentity {
    fn transform(&self, src: &[u8], dst: &mut [u8]) -> Result<(), CmsError> {
        for (g, rgb) in src.iter().zip(dst.chunks_exact_mut(3)) {
            rgb[0] = *g;
            rgb[1] = *g;
            rgb[2] = *g;
        }
        Ok(())
    }
}

impl TransformExecutor<f64> for GrayToRgbIdentity {
    fn transform(&self, src: &[f64], dst: &mut [f64]) -> Result<(), CmsError> {
        for (g, rgb) in src.iter().zip(dst.chunks_exact_mut(3)) {
            rgb[0] = *g;
            rgb[1] = *g;
            rgb[2] = *g;
        }
        Ok(())
    }
}

/// `TransformExecutor` adapter that resolves CMYK → sRGB through an existing
/// `Clut4`. Used as stage 2 of the proofing chain so the chain output goes
/// through the same hand-rolled colorimetric path as direct DeviceCMYK
/// conversion, avoiding the moxcms over-saturation cited in
/// `perceptual.rs`.
struct Clut4ToRgb {
    clut4: Clut4,
}

impl TransformExecutor<u8> for Clut4ToRgb {
    fn transform(&self, src: &[u8], dst: &mut [u8]) -> Result<(), CmsError> {
        let pixel_count = dst.len() / 3;
        let rgb = apply_clut4_cmyk_to_rgb(&self.clut4, src, pixel_count);
        dst[..rgb.len()].copy_from_slice(&rgb);
        Ok(())
    }
}

impl TransformExecutor<f64> for Clut4ToRgb {
    fn transform(&self, src: &[f64], dst: &mut [f64]) -> Result<(), CmsError> {
        let pixel_count = dst.len() / 3;
        for px in 0..pixel_count {
            let c = src[px * 4];
            let m = src[px * 4 + 1];
            let y = src[px * 4 + 2];
            let k = src[px * 4 + 3];
            let (r, g, b) = sample_clut4_single_f64(&self.clut4, c, m, y, k);
            dst[px * 3] = r;
            dst[px * 3 + 1] = g;
            dst[px * 3 + 2] = b;
        }
        Ok(())
    }
}

/// Two-stage proofing transform: `source → intermediate (OutputIntent) → sRGB`.
///
/// PDF/X workflows specify a target device via `/OutputIntents`. Source
/// colours (DeviceCMYK, ICCBased CMYK/RGB/Gray, etc.) are color-managed
/// through that target before reaching the simulated display, so every
/// source space converges through the same final `OutputIntent → sRGB`
/// stage and a CMYK swatch designed to match a DeviceCMYK background
/// renders identically to it.
///
/// `intermediate_n` is the number of components in the OutputIntent's
/// colour space (typically 4 for CMYK). The intermediate buffer is sized
/// per `transform()` call from the destination pixel count.
struct ChainedTransform<T: Copy + Default + Send + Sync + 'static> {
    /// First leg: `source → OutputIntent`.
    stage1: Arc<dyn TransformExecutor<T> + Send + Sync>,
    /// Second leg: `OutputIntent → sRGB`. Reused from the OutputIntent
    /// profile's own cached transform so byte-identical output is
    /// produced for direct DeviceCMYK paints and the proofing chain.
    stage2: Arc<dyn TransformExecutor<T> + Send + Sync>,
    /// Number of components in the intermediate (OutputIntent) layout.
    intermediate_n: usize,
}

impl<T: Copy + Default + Send + Sync + 'static> TransformExecutor<T> for ChainedTransform<T> {
    fn transform(&self, src: &[T], dst: &mut [T]) -> Result<(), CmsError> {
        // Destination is always packed sRGB (3 components). Derive pixel
        // count from `dst.len()` so we don't have to know the source
        // component count here.
        let pixel_count = dst.len() / 3;
        let mid_len = pixel_count * self.intermediate_n;
        let mut mid = vec![T::default(); mid_len];
        self.stage1.transform(src, &mut mid)?;
        self.stage2.transform(&mid, dst)
    }
}

/// Pre-baked 4D CLUT sampling a CMYK ICC transform on a regular grid.
///
/// At profile-registration time we sample moxcms at `grid_n^4` evenly-spaced
/// CMYK points and store the sRGB output. At image-conversion time we do
/// K-slice plus 3D tetrahedral interpolation inside each slice. This is ~30×
/// faster than direct moxcms for LUT-based CMYK profiles (e.g., SWOP) while
/// staying well inside imperceptible ΔE for typical print-workflow inputs.
#[derive(Clone)]
struct Clut4 {
    /// Grid points per axis (typical: 17).
    grid_n: u8,
    /// Flat LUT in order (k, y, m, c) with C fastest, K slowest.
    /// Length = grid_n^4 * 3 bytes (packed sRGB).
    data: Arc<Vec<u8>>,
}

impl Clut4 {
    /// Construct a Clut4 from a pre-baked byte buffer with the same memory
    /// layout `bake_clut4` produces (K outermost, Y, M, C innermost; 3 bytes
    /// per grid point). Used by [`perceptual::bake_clut4_perceptual`] so its
    /// output is byte-compatible with [`apply_clut4_cmyk_to_rgb`].
    fn from_baked(grid_n: u8, data: Vec<u8>) -> Self {
        debug_assert_eq!(data.len(), (grid_n as usize).pow(4) * 3);
        Self {
            grid_n,
            data: Arc::new(data),
        }
    }
}

/// Cached ICC transform to sRGB (specific to source layout).
#[derive(Clone)]
struct CachedTransform {
    /// 8-bit transform for image data. When proofing is enabled and the
    /// source is RGB, this is the Perceptual-intent chain so existing
    /// callers (no intent plumbing yet) keep producing the GWG 13.0
    /// baseline byte-for-byte.
    transform_8bit: Arc<dyn TransformExecutor<u8> + Send + Sync>,
    /// f64 transform for single-color conversions. Same intent default
    /// as `transform_8bit`.
    transform_f64: Arc<dyn TransformExecutor<f64> + Send + Sync>,
    /// Per-intent proofing chains, indexed by ICC `RenderingIntent`
    /// discriminant: `[Perceptual=0, RelCol=1, Saturation=2, AbsCol=3]`.
    /// Populated only for n=3 RGB sources when proofing is enabled and
    /// the source has a viable A2B / OI B2A pair; `None` slots fall back
    /// to `transform_*bit` / `transform_*_f64` at lookup time. AbsCol
    /// currently shares the RelCol tables (pending step 4 BPC + AbsCol
    /// white-point handling).
    chain_per_intent_8bit: [Option<Arc<dyn TransformExecutor<u8> + Send + Sync>>; 4],
    chain_per_intent_f64: [Option<Arc<dyn TransformExecutor<f64> + Send + Sync>>; 4],
    /// Per-intent stage-1 only sampler (source RGB → OutputIntent CMYK,
    /// without the trailing CMYK→sRGB stage 2). Same indexing as the
    /// `chain_per_intent_*` arrays. Used by
    /// [`IccCache::convert_to_oi_cmyk`] so the PDF reader can record the
    /// chain's intermediate CMYK as `DeviceColor::native_cmyk`, which
    /// in turn lets the renderer's `cmyk_group_blend` gate fire on
    /// ICCBased RGB swatches inside a `/CS DeviceCMYK` page group
    /// (GWG 16.1).
    chain_stage1_per_intent: [Option<Arc<perceptual::HandRolledChainStage1Rgb>>; 4],
    /// Number of source components.
    n: u32,
    /// Whether the source profile is Lab (needs value normalization).
    is_lab: bool,
    /// Pre-baked 4D CLUT for fast CMYK→sRGB image conversion.
    /// Only built for `n == 4` profiles; None otherwise.
    clut4: Option<Clut4>,
    /// Cached Black Point Compensation parameters for this profile. Computed
    /// when `n == 4` and `IccCache::bpc_mode` is enabled. Applied as a
    /// post-correction on the moxcms output (sRGB → XYZ-D50 → BPC shift →
    /// back to sRGB) so K-heavy CMYK colours map to true zero black.
    bpc_params: Option<BpcParams>,
}

/// ICC color profile cache and transform manager.
#[derive(Clone)]
pub struct IccCache {
    /// SHA-256 hash → parsed ColorProfile.
    profiles: HashMap<ProfileHash, Arc<ColorProfile>>,
    /// Cached transforms: hash → CachedTransform.
    transforms: HashMap<ProfileHash, CachedTransform>,
    /// Single-color conversion cache: `(hash-prefix, quantized_components)
    /// → (r, g, b)`. Used by [`Self::convert_color`] (the legacy /
    /// default-Perceptual path). Uses first 8 bytes of hash as u64 key
    /// for compactness.
    color_cache: HashMap<(u64, [u16; 4]), (f64, f64, f64)>,
    /// Per-intent single-color conversion cache: `(hash-prefix,
    /// quantized_components, intent_byte) → (r, g, b)`. Populated by
    /// [`Self::convert_color_with_intent`] for non-Perceptual intents
    /// so step 3's per-intent renderer plumbing can stay cheap when a
    /// page calls the same conversion repeatedly under e.g. Saturation.
    color_cache_intent: HashMap<IntentColorKey, (f64, f64, f64)>,
    /// Default system CMYK profile hash (if found at startup).
    default_cmyk_hash: Option<ProfileHash>,
    /// Raw bytes of the system CMYK profile (for re-registration in render threads).
    system_cmyk_bytes: Option<Arc<Vec<u8>>>,
    /// Raw profile bytes for each registered profile (for PDF embedding).
    raw_bytes: HashMap<ProfileHash, Arc<Vec<u8>>>,
    /// sRGB output profile (created once).
    srgb_profile: ColorProfile,
    /// Cached sRGB→CMYK reverse transform (for RGB round-trip through CMYK page groups).
    reverse_cmyk_f64: Option<Arc<dyn TransformExecutor<f64> + Send + Sync>>,
    /// Black Point Compensation mode for CMYK→sRGB conversion. Set at
    /// construction time via [`IccCacheOptions`]; consulted by future BPC
    /// apply paths (commit 2 of `docs/PLAN-BPC.md`).
    bpc_mode: BpcMode,
    /// Enable PDF/X-style proofing: ICCBased source profiles convert through
    /// the default CMYK profile (the document's OutputIntent) before going
    /// to sRGB, so all source colour spaces converge through the same
    /// `OutputIntent → sRGB` final stage. Set by
    /// `PdfDocument::apply_output_intent_as_default_cmyk` when the PDF has
    /// an `/OutputIntents` entry; left `false` when only the system CMYK
    /// fallback is available, so non-PDF/X documents keep their direct
    /// `source → sRGB` conversion.
    proofing_enabled: bool,
    /// Per-intent Lab → OutputIntent CMYK samplers. Built lazily by
    /// [`Self::prepare_lab_to_oi_cmyk`] from the OI's `B2A*` LUTs so
    /// `convert_lab_to_oi_cmyk` can populate `DeviceColor::native_cmyk`
    /// for Lab fills with a direct `Lab → PCS → OI B2A → CMYK` value
    /// (matching Acrobat's ACE) instead of the sRGB→ICC-reverse approximation
    /// the parallel CMYK buffer would otherwise compute. Surfaced under
    /// CMYK-group blends — GWG 22.1's ColorBurn form over a Lab BG is the
    /// canonical case where the indirect path drifts visibly.
    lab_to_oi_per_intent: [Option<Arc<perceptual::LabToCmykSampler>>; 4],
}

impl Default for IccCache {
    fn default() -> Self {
        Self::new()
    }
}

/// Apply BPC to an sRGB triple if `params` is `Some`; otherwise return the
/// triple unchanged. Centralised so every conversion entry point stays in
/// sync.
#[inline]
fn bpc_post_correct(rgb: [f64; 3], params: Option<&BpcParams>) -> [f64; 3] {
    match params {
        Some(p) => apply_bpc_f64(rgb, p),
        None => rgb,
    }
}

impl IccCache {
    /// Create an empty ICC cache with default options (BPC `Auto`, no
    /// pre-supplied source CMYK profile).
    pub fn new() -> Self {
        Self::new_with_options(IccCacheOptions::default())
    }

    /// Create an ICC cache with the given options.
    ///
    /// When `opts.source_cmyk_profile` is `Some`, the bytes are registered as
    /// the system CMYK profile (overriding any later
    /// [`Self::search_system_cmyk_profile`] call). Otherwise the cache starts
    /// empty and the caller is expected to supply a profile separately.
    pub fn new_with_options(opts: IccCacheOptions) -> Self {
        let mut cache = Self {
            profiles: HashMap::new(),
            transforms: HashMap::new(),
            color_cache: HashMap::new(),
            color_cache_intent: HashMap::new(),
            default_cmyk_hash: None,
            system_cmyk_bytes: None,
            raw_bytes: HashMap::new(),
            srgb_profile: ColorProfile::new_srgb(),
            reverse_cmyk_f64: None,
            bpc_mode: opts.bpc_mode,
            proofing_enabled: false,
            lab_to_oi_per_intent: [None, None, None, None],
        };
        if let Some(bytes) = opts.source_cmyk_profile {
            cache.load_cmyk_profile_bytes(&bytes);
        }
        cache
    }

    /// Current Black Point Compensation mode.
    #[inline]
    pub fn bpc_mode(&self) -> BpcMode {
        self.bpc_mode
    }

    /// Compute the SHA-256 hash of an ICC profile without registering it.
    pub fn hash_profile(bytes: &[u8]) -> ProfileHash {
        use sha2::{Digest, Sha256};
        Sha256::digest(bytes).into()
    }

    /// Register an ICC profile from raw bytes. Returns the SHA-256 hash on success.
    pub fn register_profile(&mut self, bytes: &[u8]) -> Option<ProfileHash> {
        self.register_profile_with_n(bytes, None)
    }

    /// Register an ICC profile, validating that its color space matches the
    /// expected component count `expected_n`. When the profile's actual color
    /// space has a different number of components (e.g. an RGB profile stored
    /// with PDF `/N 1`), the profile is rejected so the caller can fall back
    /// to the alternate color space.
    pub fn register_profile_with_n(
        &mut self,
        bytes: &[u8],
        expected_n: Option<u32>,
    ) -> Option<ProfileHash> {
        use sha2::{Digest, Sha256};
        let hash: ProfileHash = Sha256::digest(bytes).into();

        // Already registered?
        if self.transforms.contains_key(&hash) {
            return Some(hash);
        }

        // Store raw bytes for PDF embedding
        self.raw_bytes
            .entry(hash)
            .or_insert_with(|| Arc::new(bytes.to_vec()));

        let profile = match ColorProfile::new_from_slice(bytes) {
            Ok(p) => p,
            Err(e) => {
                eprintln!("[ICC] Failed to parse profile: {e}");
                return None;
            }
        };

        let n = match profile.color_space {
            DataColorSpace::Gray => 1u32,
            DataColorSpace::Rgb => 3,
            DataColorSpace::Cmyk => 4,
            DataColorSpace::Lab => 3,
            _ => {
                eprintln!(
                    "[ICC] Unsupported profile color space: {:?}",
                    profile.color_space
                );
                return None;
            }
        };

        // Reject profile when its actual component count doesn't match the
        // PDF's /N declaration — the input data won't match the profile's
        // expected input layout.
        if let Some(expected) = expected_n {
            if n != expected {
                return None;
            }
        }

        let (src_layout_8, src_layout_f64) = match n {
            1 => (Layout::Gray, Layout::Gray),
            3 => (Layout::Rgb, Layout::Rgb),
            4 => (Layout::Rgba, Layout::Rgba),
            _ => return None,
        };

        let dst_layout_8 = Layout::Rgb;
        let dst_layout_f64 = Layout::Rgb;

        // Try multiple rendering intents — Perceptual first so the
        // moxcms-driven `transform_f64` Arc honours the profile's perceptual
        // table. Most CMYK paths route through the perceptual A2B0 CLUT
        // (`bake_clut4_perceptual`) instead, but a few sites still call the
        // f64 transform directly (e.g. `Luminosity` soft-mask conversion in
        // the renderer); for those sites, picking the Perceptual transform
        // here keeps the per-pixel result aligned with the CLUT path. ICC v4
        // profiles may only have A2B0, so this also covers those.
        let intents = [
            RenderingIntent::Perceptual,
            RenderingIntent::RelativeColorimetric,
            RenderingIntent::AbsoluteColorimetric,
            RenderingIntent::Saturation,
        ];

        let mut transform_8bit = None;
        for &intent in &intents {
            let options = TransformOptions {
                rendering_intent: intent,
                ..TransformOptions::default()
            };
            match profile.create_transform_8bit(
                src_layout_8,
                &self.srgb_profile,
                dst_layout_8,
                options,
            ) {
                Ok(t) => {
                    transform_8bit = Some(t);
                    break;
                }
                Err(_) => continue,
            }
        }
        let transform_8bit = match transform_8bit {
            Some(t) => t,
            None if n == 1 => {
                // Gray profiles that can't produce Gray→sRGB transforms (e.g.
                // minimal Linotype profiles with only a TRC): fall back to the
                // sRGB gray curve, which is functionally correct for most Gray
                // profiles encountered in PDFs.
                return self.register_gray_identity(hash, profile);
            }
            None => {
                eprintln!(
                    "[ICC] Failed to create 8-bit transform (cs={:?})",
                    profile.color_space
                );
                return None;
            }
        };

        let mut transform_f64 = None;
        for &intent in &intents {
            let options = TransformOptions {
                rendering_intent: intent,
                ..TransformOptions::default()
            };
            match profile.create_transform_f64(
                src_layout_f64,
                &self.srgb_profile,
                dst_layout_f64,
                options,
            ) {
                Ok(t) => {
                    transform_f64 = Some(t);
                    break;
                }
                Err(_) => continue,
            }
        }
        let transform_f64 = match transform_f64 {
            Some(t) => t,
            None if n == 1 => {
                // Same Gray fallback for f64 path
                return self.register_gray_identity(hash, profile);
            }
            None => {
                eprintln!(
                    "[ICC] Failed to create f64 transform (cs={:?})",
                    profile.color_space
                );
                return None;
            }
        };

        let is_lab = profile.color_space == DataColorSpace::Lab;

        // PDF/X proofing chain: when proofing is enabled and a non-OutputIntent
        // profile is being registered, route source colours through the
        // OutputIntent (`source → OutputIntent → sRGB`) so that an ICCBased
        // CMYK swatch designed to match a DeviceCMYK background renders
        // identically to it. The OutputIntent itself is registered with a
        // direct transform; subsequent profiles are chained.
        type ChainPair = (
            Arc<dyn TransformExecutor<u8> + Send + Sync>,
            Arc<dyn TransformExecutor<f64> + Send + Sync>,
        );
        let mut chain_per_intent_8bit: [Option<Arc<dyn TransformExecutor<u8> + Send + Sync>>; 4] =
            Default::default();
        let mut chain_per_intent_f64: [Option<Arc<dyn TransformExecutor<f64> + Send + Sync>>; 4] =
            Default::default();
        let mut chain_stage1_per_intent: [Option<Arc<perceptual::HandRolledChainStage1Rgb>>; 4] =
            Default::default();
        let chain_data: Option<ChainPair> = if self.proofing_enabled
            && let Some(oi_hash) = self.default_cmyk_hash
            && oi_hash != hash
            && let Some(oi_profile) = self.profiles.get(&oi_hash).cloned()
            && let Some(oi_cached) = self.transforms.get(&oi_hash)
            && oi_cached.n == 4
        {
            // OI is always CMYK in PDF/X workflows; intermediate is 4-component.
            let oi_layout_8 = Layout::Rgba;
            let oi_layout_f64 = Layout::Rgba;

            // Stage 2 (`OutputIntent → sRGB`): prefer OI's hand-rolled
            // CLUT4 over its moxcms transform so chain output matches
            // direct DeviceCMYK conversion at the byte level. The same
            // stage-2 is reused across every intent's chain.
            let stage2_8bit: Arc<dyn TransformExecutor<u8> + Send + Sync> =
                if let Some(oi_clut) = oi_cached.clut4.clone() {
                    Arc::new(Clut4ToRgb { clut4: oi_clut })
                } else {
                    oi_cached.transform_8bit.clone()
                };
            let stage2_f64: Arc<dyn TransformExecutor<f64> + Send + Sync> =
                if let Some(oi_clut) = oi_cached.clut4.clone() {
                    Arc::new(Clut4ToRgb { clut4: oi_clut })
                } else {
                    oi_cached.transform_f64.clone()
                };

            // Hand-rolled chain stage 1 for RGB sources. moxcms's
            // `create_transform` over-saturates sRGB-style first legs by
            // ~5–15% (visible as the GWG 16.1 "X" marks), so we compose
            // `source.A2B[i] → OI.B2A[i]` per-pixel using the same
            // primitives moxcms exposes. Per-pixel composition (vs. an
            // intermediate 17³ CLUT bake) avoids the quantization layer
            // that drifts GWG 13.0's BG-vs-X match by ~6 RGB levels. Only
            // n=3 (RGB) sources go through this path; CMYK sources keep
            // using moxcms (no diagnostic shows that leg drifting).
            //
            // For each of the four ICC rendering intents we try to build
            // a chain using that intent's tables on both sides. Intents
            // whose A2B / B2A tables are missing (or whose profile is in
            // an unsupported shape) skip silently and the slot stays
            // `None`; lookup-time callers fall back to moxcms.
            if n == 3 {
                use moxcms::RenderingIntent;
                for &intent in &[
                    RenderingIntent::Perceptual,
                    RenderingIntent::RelativeColorimetric,
                    RenderingIntent::Saturation,
                    RenderingIntent::AbsoluteColorimetric,
                ] {
                    let Some(stage1) =
                        perceptual::HandRolledChainStage1Rgb::new(&profile, &oi_profile, intent)
                    else {
                        continue;
                    };
                    let stage1_arc = Arc::new(stage1);
                    let chain_8: Arc<dyn TransformExecutor<u8> + Send + Sync> =
                        Arc::new(ChainedTransform {
                            stage1: stage1_arc.clone(),
                            stage2: stage2_8bit.clone(),
                            intermediate_n: 4,
                        });
                    let chain_f: Arc<dyn TransformExecutor<f64> + Send + Sync> =
                        Arc::new(ChainedTransform {
                            stage1: stage1_arc.clone(),
                            stage2: stage2_f64.clone(),
                            intermediate_n: 4,
                        });
                    let i = intent as usize;
                    chain_per_intent_8bit[i] = Some(chain_8);
                    chain_per_intent_f64[i] = Some(chain_f);
                    chain_stage1_per_intent[i] = Some(stage1_arc);
                }
            }

            // Default chain (used when the lookup path doesn't yet pass an
            // intent — the current state of `convert_color`). Picks the
            // Perceptual hand-rolled chain when available; otherwise falls
            // back to moxcms's transform-driven build, preserving the
            // n=4 CMYK source path.
            let perceptual_idx = moxcms::RenderingIntent::Perceptual as usize;
            if let (Some(c8), Some(cf)) = (
                chain_per_intent_8bit[perceptual_idx].clone(),
                chain_per_intent_f64[perceptual_idx].clone(),
            ) {
                Some((c8, cf))
            } else {
                let mut stage1_8bit_opt: Option<Arc<dyn TransformExecutor<u8> + Send + Sync>> =
                    None;
                let mut stage1_f64_opt: Option<Arc<dyn TransformExecutor<f64> + Send + Sync>> =
                    None;
                for &intent in &intents {
                    let options = TransformOptions {
                        rendering_intent: intent,
                        ..TransformOptions::default()
                    };
                    if let Ok(t) = profile.create_transform_8bit(
                        src_layout_8,
                        &oi_profile,
                        oi_layout_8,
                        options,
                    ) {
                        stage1_8bit_opt = Some(t);
                        break;
                    }
                }
                for &intent in &intents {
                    let options = TransformOptions {
                        rendering_intent: intent,
                        ..TransformOptions::default()
                    };
                    if let Ok(t) = profile.create_transform_f64(
                        src_layout_f64,
                        &oi_profile,
                        oi_layout_f64,
                        options,
                    ) {
                        stage1_f64_opt = Some(t);
                        break;
                    }
                }
                match (stage1_8bit_opt, stage1_f64_opt) {
                    (Some(s1_8), Some(s1_f)) => {
                        let chain_8: Arc<dyn TransformExecutor<u8> + Send + Sync> =
                            Arc::new(ChainedTransform {
                                stage1: s1_8,
                                stage2: stage2_8bit,
                                intermediate_n: 4,
                            });
                        let chain_f: Arc<dyn TransformExecutor<f64> + Send + Sync> =
                            Arc::new(ChainedTransform {
                                stage1: s1_f,
                                stage2: stage2_f64,
                                intermediate_n: 4,
                            });
                        Some((chain_8, chain_f))
                    }
                    _ => None,
                }
            }
        } else {
            None
        };

        // When the chain build succeeded, override the cached transforms with
        // the chain versions. The CLUT4 below is then baked from the chain
        // (for CMYK source profiles), so single-color and image conversions
        // share the same `OutputIntent → sRGB` final stage as DeviceCMYK.
        let (transform_8bit, transform_f64, chain_active) = match chain_data {
            Some((c8, cf)) => (c8, cf, true),
            None => (transform_8bit, transform_f64, false),
        };

        // For 4-channel (CMYK) profiles, pre-bake a 17^4 CLUT for fast image
        // conversion. Two paths produce the same Clut4 layout:
        //
        // 1. `bake_clut4_perceptual` samples the profile's own A2B1
        //    (colorimetric) table directly, decodes the legacy v2 PCS-Lab
        //    encoding, and clips out-of-gamut colours to the sRGB boundary.
        //    Output matches lcms2's `cmsDoTransform(RelCol)` to ±1 RGB level.
        //    Available for v2 mft2 CMYK profiles. BPC is computed inside the
        //    bake against this sampler's own (1,1,1,1) output so the source
        //    black-point matches what we're actually producing.
        // 2. `bake_clut4` invokes the 8-bit moxcms transform on a grid;
        //    fallback for profiles whose tables are missing or in a shape we
        //    don't yet handle (mAB, mft1, XYZ-PCS). BPC is calibrated against
        //    moxcms's transform output (`detect_source_black_point`).
        //
        // The runtime CLUT lookup is identical regardless of which path
        // produced the table.
        //
        // `bpc_params` is stored on `CachedTransform` for the moxcms-fallback
        // path's `convert_color` / `convert_color_readonly` callers when the
        // bake returned `None`. The hand-rolled sampler folds BPC in directly
        // and leaves this `None`; the cached `transform_f64` Arc is only
        // exercised in fallback contexts and shouldn't double-apply BPC.
        let bpc_enabled = n == 4 && self.bpc_mode.is_enabled();
        let mut bpc_params: Option<BpcParams> = None;
        let clut4 = if n == 4 && !chain_active {
            // Direct (non-proofing) CMYK profiles: pre-bake a CLUT for fast
            // image conversion.
            let c = perceptual::bake_clut4_perceptual(&profile, 17, bpc_enabled).or_else(|| {
                let params = if bpc_enabled {
                    detect_source_black_point(transform_8bit.as_ref())
                        .map(|sbp| compute_bpc_params(sbp, [0.0; 3], bpc::WP_D50))
                } else {
                    None
                };
                let r = bake_clut4(transform_8bit.as_ref(), 17, params.as_ref());
                bpc_params = params;
                r
            });
            if std::env::var_os("STET_ICC_VERIFY").is_some()
                && let Some(ref clut) = c
            {
                verify_clut4(clut, transform_8bit.as_ref(), bpc_params.as_ref());
            }
            c
        } else {
            // Chain-mode: do NOT pre-bake the source profile's CLUT. A
            // pre-baked source CLUT would compose two CLUT4 quantizations
            // (source-bake + stage 2's OutputIntent-bake) and drift a few
            // sRGB levels relative to direct DeviceCMYK paints. Routing every
            // call through the chain transform keeps just one CLUT
            // quantization (OutputIntent's, in stage 2), so a designed-equal
            // ICCBased CMYK swatch and DeviceCMYK background converge.
            None
        };

        self.profiles.insert(hash, Arc::new(profile));
        self.transforms.insert(
            hash,
            CachedTransform {
                transform_8bit,
                transform_f64,
                chain_per_intent_8bit,
                chain_per_intent_f64,
                chain_stage1_per_intent,
                n,
                is_lab,
                clut4,
                bpc_params,
            },
        );

        Some(hash)
    }

    /// Register a Gray profile with an identity Gray→RGB fallback transform.
    /// Used when the ICC library can't create a proper transform from the profile
    /// (e.g. minimal profiles with only a TRC and no A2B/B2A tables).
    fn register_gray_identity(
        &mut self,
        hash: ProfileHash,
        profile: ColorProfile,
    ) -> Option<ProfileHash> {
        self.profiles.insert(hash, Arc::new(profile));
        self.transforms.insert(
            hash,
            CachedTransform {
                transform_8bit: Arc::new(GrayToRgbIdentity),
                transform_f64: Arc::new(GrayToRgbIdentity),
                chain_per_intent_8bit: Default::default(),
                chain_per_intent_f64: Default::default(),
                chain_stage1_per_intent: Default::default(),
                n: 1,
                is_lab: false,
                clut4: None,
                bpc_params: None,
            },
        );
        Some(hash)
    }

    /// Convert RGB components through the proofing chain's stage 1 to
    /// the OutputIntent's CMYK ink values. Returns `None` when no
    /// hand-rolled chain exists for this profile (the document isn't
    /// PDF/X, or the profile shape is unsupported), in which case the
    /// caller should leave `DeviceColor::native_cmyk` as `None` and the
    /// renderer falls back to its sRGB-derived approximation. The
    /// `intent` parameter selects which per-intent chain to use; falls
    /// back to the Perceptual chain when the requested intent slot is
    /// empty.
    pub fn convert_to_oi_cmyk(
        &self,
        hash: &ProfileHash,
        components: &[f64],
        intent: RenderingIntent,
    ) -> Option<[f64; 4]> {
        let cached = self.transforms.get(hash)?;
        if cached.n != 3 {
            return None;
        }
        let stage1 = cached.chain_stage1_per_intent[intent as usize]
            .as_ref()
            .or(cached.chain_stage1_per_intent[RenderingIntent::Perceptual as usize].as_ref())?;
        let r = components.first().copied().unwrap_or(0.0).clamp(0.0, 1.0);
        let g = components.get(1).copied().unwrap_or(0.0).clamp(0.0, 1.0);
        let b = components.get(2).copied().unwrap_or(0.0).clamp(0.0, 1.0);
        Some(stage1.sample_cmyk_f64(r, g, b))
    }

    /// Convert a single color through an ICC profile to sRGB.
    /// Returns (r, g, b) in [0, 1] range.
    pub fn convert_color(
        &mut self,
        hash: &ProfileHash,
        components: &[f64],
    ) -> Option<(f64, f64, f64)> {
        let cached = self.transforms.get(hash)?;
        let n = cached.n as usize;
        let is_lab = cached.is_lab;

        // Normalize input values to [0,1] range.
        // Lab profiles need special mapping: L/100, (a+128)/255, (b+128)/255.
        let mut src = vec![0.0f64; n];
        for (i, s) in src.iter_mut().enumerate() {
            let v = components.get(i).copied().unwrap_or(0.0);
            *s = if is_lab {
                match i {
                    0 => (v / 100.0).clamp(0.0, 1.0),
                    _ => ((v + 128.0) / 255.0).clamp(0.0, 1.0),
                }
            } else {
                v.clamp(0.0, 1.0)
            };
        }

        // Quantize normalized values for cache key
        let hash_prefix = u64::from_le_bytes(hash[..8].try_into().ok()?);
        let mut quantized = [0u16; 4];
        for (i, &c) in src.iter().take(4).enumerate() {
            quantized[i] = (c * 65535.0).round() as u16;
        }

        // Check cache
        let cache_key = (hash_prefix, quantized);
        if let Some(&cached) = self.color_cache.get(&cache_key) {
            return Some(cached);
        }

        let result = if n == 4
            && let Some(clut) = cached.clut4.as_ref()
        {
            // Route single-color CMYK through the same baked CLUT image
            // conversions use, so a flat fill matches the surrounding gradient
            // stops byte-for-byte. BPC and the perceptual A2B0 sampling are
            // already folded into the CLUT.
            let (r, g, b) = sample_clut4_single_f64(clut, src[0], src[1], src[2], src[3]);
            (r, g, b)
        } else {
            let mut dst = [0.0f64; 3];
            if cached.transform_f64.transform(&src, &mut dst).is_err() {
                return None;
            }
            let dst = bpc_post_correct(dst, cached.bpc_params.as_ref());
            (
                dst[0].clamp(0.0, 1.0),
                dst[1].clamp(0.0, 1.0),
                dst[2].clamp(0.0, 1.0),
            )
        };

        // Cache (limit size to avoid unbounded growth)
        if self.color_cache.len() < 65536 {
            self.color_cache.insert(cache_key, result);
        }

        Some(result)
    }

    /// Cached single-color conversion under a specific rendering
    /// intent. Delegates to the legacy [`Self::convert_color`] for
    /// Perceptual (which uses the Perceptual chain stored in
    /// `transform_f64` and the Perceptual-keyed cache); for the other
    /// intents it goes through [`Self::convert_color_readonly_with_intent`]
    /// and caches per-intent.
    pub fn convert_color_with_intent(
        &mut self,
        hash: &ProfileHash,
        components: &[f64],
        intent: RenderingIntent,
    ) -> Option<(f64, f64, f64)> {
        if matches!(intent, RenderingIntent::Perceptual) {
            return self.convert_color(hash, components);
        }
        // Quantize for cache key. Have to recompute here because
        // `convert_color_readonly_with_intent` doesn't return the
        // quantized buffer. The arithmetic mirrors `convert_color`.
        let cached = self.transforms.get(hash)?;
        let n = cached.n as usize;
        let is_lab = cached.is_lab;
        let mut src = vec![0.0f64; n];
        for (i, s) in src.iter_mut().enumerate() {
            let v = components.get(i).copied().unwrap_or(0.0);
            *s = if is_lab {
                match i {
                    0 => (v / 100.0).clamp(0.0, 1.0),
                    _ => ((v + 128.0) / 255.0).clamp(0.0, 1.0),
                }
            } else {
                v.clamp(0.0, 1.0)
            };
        }
        let hash_prefix = u64::from_le_bytes(hash[..8].try_into().ok()?);
        let mut quantized = [0u16; 4];
        for (i, &c) in src.iter().take(4).enumerate() {
            quantized[i] = (c * 65535.0).round() as u16;
        }
        let cache_key = (hash_prefix, quantized, intent as u8);
        if let Some(&hit) = self.color_cache_intent.get(&cache_key) {
            return Some(hit);
        }
        let result = self.convert_color_readonly_with_intent(hash, components, intent)?;
        if self.color_cache_intent.len() < 65536 {
            self.color_cache_intent.insert(cache_key, result);
        }
        Some(result)
    }

    /// Convert a single color through an ICC profile using a specific
    /// rendering intent. Falls back to the cached default chain (built
    /// from the Perceptual tables) when no per-intent chain is
    /// available — that path matches [`Self::convert_color_readonly`]
    /// byte-for-byte and is the common case for non-PDF/X documents and
    /// CMYK source profiles.
    pub fn convert_color_readonly_with_intent(
        &self,
        hash: &ProfileHash,
        components: &[f64],
        intent: RenderingIntent,
    ) -> Option<(f64, f64, f64)> {
        let cached = self.transforms.get(hash)?;
        let n = cached.n as usize;
        let is_lab = cached.is_lab;

        let mut src = vec![0.0f64; n];
        for (i, s) in src.iter_mut().enumerate() {
            let v = components.get(i).copied().unwrap_or(0.0);
            *s = if is_lab {
                match i {
                    0 => (v / 100.0).clamp(0.0, 1.0),
                    _ => ((v + 128.0) / 255.0).clamp(0.0, 1.0),
                }
            } else {
                v.clamp(0.0, 1.0)
            };
        }

        // CMYK profiles: route through the pre-baked CLUT4 (no per-intent
        // variant exists on this path; intent-driven CMYK B2A selection
        // is deferred until the chain-side per-intent OI bake).
        if n == 4
            && let Some(clut) = cached.clut4.as_ref()
        {
            let (r, g, b) = sample_clut4_single_f64(clut, src[0], src[1], src[2], src[3]);
            return Some((r, g, b));
        }

        let transform = cached.chain_per_intent_f64[intent as usize]
            .as_ref()
            .unwrap_or(&cached.transform_f64);
        let mut dst = [0.0f64; 3];
        if transform.transform(&src, &mut dst).is_err() {
            return None;
        }

        let dst = bpc_post_correct(dst, cached.bpc_params.as_ref());
        Some((
            dst[0].clamp(0.0, 1.0),
            dst[1].clamp(0.0, 1.0),
            dst[2].clamp(0.0, 1.0),
        ))
    }

    /// Convert a single color through an ICC profile (read-only, no caching).
    ///
    /// Same as `convert_color` but takes `&self` instead of `&mut self`,
    /// suitable for use from immutable contexts like rendering.
    pub fn convert_color_readonly(
        &self,
        hash: &ProfileHash,
        components: &[f64],
    ) -> Option<(f64, f64, f64)> {
        let cached = self.transforms.get(hash)?;
        let n = cached.n as usize;
        let is_lab = cached.is_lab;

        let mut src = vec![0.0f64; n];
        for (i, s) in src.iter_mut().enumerate() {
            let v = components.get(i).copied().unwrap_or(0.0);
            *s = if is_lab {
                match i {
                    0 => (v / 100.0).clamp(0.0, 1.0),
                    _ => ((v + 128.0) / 255.0).clamp(0.0, 1.0),
                }
            } else {
                v.clamp(0.0, 1.0)
            };
        }

        if n == 4
            && let Some(clut) = cached.clut4.as_ref()
        {
            let (r, g, b) = sample_clut4_single_f64(clut, src[0], src[1], src[2], src[3]);
            return Some((r, g, b));
        }

        let mut dst = [0.0f64; 3];
        if cached.transform_f64.transform(&src, &mut dst).is_err() {
            return None;
        }

        let dst = bpc_post_correct(dst, cached.bpc_params.as_ref());
        Some((
            dst[0].clamp(0.0, 1.0),
            dst[1].clamp(0.0, 1.0),
            dst[2].clamp(0.0, 1.0),
        ))
    }

    /// Bulk-convert 8-bit image samples through an ICC profile to RGB
    /// using a specific rendering intent. Falls back to the cached
    /// default 8-bit transform (built from the Perceptual tables) when
    /// no per-intent chain is available — that path matches
    /// [`Self::convert_image_8bit`] byte-for-byte.
    pub fn convert_image_8bit_with_intent(
        &self,
        hash: &ProfileHash,
        samples: &[u8],
        pixel_count: usize,
        intent: RenderingIntent,
    ) -> Option<Vec<u8>> {
        let cached = self.transforms.get(hash)?;
        let n = cached.n as usize;
        let expected_len = pixel_count * n;
        if samples.len() < expected_len {
            return None;
        }

        // CMYK profiles route through the pre-baked CLUT4. Per-intent
        // CMYK B2A selection on this side of the chain is deferred —
        // step 4 will revisit when BPC + AbsCol land.
        if let Some(clut) = &cached.clut4 {
            return Some(apply_clut4_cmyk_to_rgb(
                clut,
                &samples[..expected_len],
                pixel_count,
            ));
        }

        let transform = cached.chain_per_intent_8bit[intent as usize]
            .as_ref()
            .unwrap_or(&cached.transform_8bit);
        let src = &samples[..expected_len];
        let mut dst = vec![0u8; pixel_count * 3];
        match transform.transform(src, &mut dst) {
            Ok(()) => {
                if let Some(p) = cached.bpc_params.as_ref() {
                    for px in dst.chunks_exact_mut(3) {
                        let out = apply_bpc_rgb_u8([px[0], px[1], px[2]], p);
                        px[0] = out[0];
                        px[1] = out[1];
                        px[2] = out[2];
                    }
                }
                Some(dst)
            }
            Err(e) => {
                eprintln!("[ICC] Image transform failed (intent {intent:?}): {e}");
                None
            }
        }
    }

    /// Bulk-convert 8-bit image samples through an ICC profile to RGB.
    /// Input: packed samples (Gray/RGB/CMYK depending on profile).
    /// Output: packed RGB bytes (3 bytes per pixel).
    pub fn convert_image_8bit(
        &self,
        hash: &ProfileHash,
        samples: &[u8],
        pixel_count: usize,
    ) -> Option<Vec<u8>> {
        let cached = self.transforms.get(hash)?;
        let n = cached.n as usize;
        let expected_len = pixel_count * n;
        if samples.len() < expected_len {
            return None;
        }

        // Fast path: pre-baked 4D CLUT for CMYK profiles. BPC is already
        // baked into the CLUT (when enabled), so no per-pixel correction
        // is needed here.
        if let Some(clut) = &cached.clut4 {
            return Some(apply_clut4_cmyk_to_rgb(
                clut,
                &samples[..expected_len],
                pixel_count,
            ));
        }

        let src = &samples[..expected_len];
        let mut dst = vec![0u8; pixel_count * 3];

        match cached.transform_8bit.transform(src, &mut dst) {
            Ok(()) => {
                // Apply BPC per pixel for non-CLUT bulk paths (CMYK profiles
                // whose CLUT bake failed; today no other layouts populate
                // bpc_params, so this is a no-op for RGB/Gray/Lab).
                if let Some(p) = cached.bpc_params.as_ref() {
                    for px in dst.chunks_exact_mut(3) {
                        let out = apply_bpc_rgb_u8([px[0], px[1], px[2]], p);
                        px[0] = out[0];
                        px[1] = out[1];
                        px[2] = out[2];
                    }
                }
                Some(dst)
            }
            Err(e) => {
                eprintln!("[ICC] Image transform failed: {e}");
                None
            }
        }
    }

    /// Search system paths for a CMYK ICC profile and register it.
    pub fn search_system_cmyk_profile(&mut self) {
        if let Some(bytes) = find_system_cmyk_profile()
            && let Some(hash) = self.register_profile(&bytes)
        {
            eprintln!("[ICC] Loaded system CMYK profile");
            self.system_cmyk_bytes = Some(Arc::new(bytes));
            self.default_cmyk_hash = Some(hash);
        }
    }

    /// Load a CMYK ICC profile from raw bytes (for environments without filesystem access).
    pub fn load_cmyk_profile_bytes(&mut self, bytes: &[u8]) {
        if let Some(hash) = self.register_profile(bytes) {
            self.system_cmyk_bytes = Some(Arc::new(bytes.to_vec()));
            self.default_cmyk_hash = Some(hash);
        }
    }

    /// Get the default CMYK profile hash, if a system CMYK profile was found.
    pub fn default_cmyk_hash(&self) -> Option<&ProfileHash> {
        self.default_cmyk_hash.as_ref()
    }

    /// Check if a profile hash has been registered.
    pub fn has_profile(&self, hash: &ProfileHash) -> bool {
        self.transforms.contains_key(hash)
    }

    /// Get the raw bytes of a registered ICC profile (for PDF embedding).
    pub fn get_profile_bytes(&self, hash: &ProfileHash) -> Option<Arc<Vec<u8>>> {
        self.raw_bytes.get(hash).cloned()
    }

    /// Get the raw bytes of the system CMYK profile (for re-registration in render threads).
    pub fn system_cmyk_bytes(&self) -> Option<&Arc<Vec<u8>>> {
        self.system_cmyk_bytes.as_ref()
    }

    /// Set the system CMYK profile from pre-loaded bytes and hash.
    ///
    /// Used by `--output-profile` to substitute the auto-detected system CMYK
    /// profile with a user-specified one.
    pub fn set_system_cmyk(&mut self, bytes: &[u8], hash: ProfileHash) {
        self.system_cmyk_bytes = Some(Arc::new(bytes.to_vec()));
        self.default_cmyk_hash = Some(hash);
    }

    /// Enable PDF/X-style proofing: ICCBased profiles registered while this
    /// flag is set route through the default CMYK profile (the OutputIntent)
    /// before reaching sRGB. See [`Self::proofing_enabled`].
    pub fn set_proofing_enabled(&mut self, enabled: bool) {
        self.proofing_enabled = enabled;
    }

    /// Whether PDF/X-style proofing is enabled. When `true`, source ICC
    /// profiles other than the default CMYK convert through the default
    /// CMYK ("OutputIntent") instead of going directly to sRGB. This makes
    /// per-swatch ICC colours and surrounding DeviceCMYK paints converge
    /// through the same final `OutputIntent → sRGB` stage. When `false`,
    /// every profile converts directly (the historical behaviour).
    pub fn proofing_enabled(&self) -> bool {
        self.proofing_enabled
    }

    /// Set the default CMYK profile hash (used when building render-thread caches).
    pub fn set_default_cmyk_hash(&mut self, hash: ProfileHash) {
        self.default_cmyk_hash = Some(hash);
    }

    /// Temporarily remove the default CMYK hash, returning the old value.
    /// Used to disable ICC CMYK conversion inside soft mask form rendering,
    /// where PLRM formulas produce correct luminosity values (ICC profiles
    /// map 100% K to non-zero RGB, breaking luminosity soft masks).
    pub fn suspend_default_cmyk(&mut self) -> Option<ProfileHash> {
        self.default_cmyk_hash.take()
    }

    /// Restore a previously suspended default CMYK hash.
    pub fn restore_default_cmyk(&mut self, hash: Option<ProfileHash>) {
        self.default_cmyk_hash = hash;
    }

    /// Convert CMYK to (r, g, b) using the default system CMYK profile.
    /// Returns None if no system CMYK profile is loaded.
    #[inline]
    pub fn convert_cmyk(&mut self, c: f64, m: f64, y: f64, k: f64) -> Option<(f64, f64, f64)> {
        let hash = *self.default_cmyk_hash.as_ref()?;
        self.convert_color(&hash, &[c, m, y, k])
    }

    /// Convert CMYK to (r, g, b) using the default system CMYK profile (read-only, no caching).
    /// Used by band renderers that only have `&self` access.
    pub fn convert_cmyk_readonly(&self, c: f64, m: f64, y: f64, k: f64) -> Option<(f64, f64, f64)> {
        let hash = self.default_cmyk_hash.as_ref()?;
        let cached = self.transforms.get(hash)?;
        let src = [
            c.clamp(0.0, 1.0),
            m.clamp(0.0, 1.0),
            y.clamp(0.0, 1.0),
            k.clamp(0.0, 1.0),
        ];
        if let Some(clut) = cached.clut4.as_ref() {
            // Sample the same baked CLUT image conversions use, so a flat fill
            // matches the surrounding gradient stops byte-for-byte. BPC and the
            // perceptual A2B0 sampling are already folded into the CLUT.
            let (r, g, b) = sample_clut4_single_f64(clut, src[0], src[1], src[2], src[3]);
            return Some((r, g, b));
        }
        let mut dst = [0.0f64; 3];
        if cached.transform_f64.transform(&src, &mut dst).is_err() {
            return None;
        }
        let dst = bpc_post_correct(dst, cached.bpc_params.as_ref());
        Some((
            dst[0].clamp(0.0, 1.0),
            dst[1].clamp(0.0, 1.0),
            dst[2].clamp(0.0, 1.0),
        ))
    }

    /// Build the lazy sRGB→CMYK reverse transform from the system CMYK profile.
    /// Returns `Some(())` if the transform is now present (built or already
    /// cached). Returns `None` if no system CMYK profile is registered or no
    /// rendering intent could create a transform.
    fn ensure_reverse_cmyk_transform(&mut self) -> Option<()> {
        if self.reverse_cmyk_f64.is_some() {
            return Some(());
        }
        let hash = *self.default_cmyk_hash.as_ref()?;
        let cmyk_profile = self.profiles.get(&hash)?.clone();
        let intents = [
            RenderingIntent::RelativeColorimetric,
            RenderingIntent::Perceptual,
            RenderingIntent::AbsoluteColorimetric,
            RenderingIntent::Saturation,
        ];
        for &intent in &intents {
            let options = TransformOptions {
                rendering_intent: intent,
                ..TransformOptions::default()
            };
            if let Ok(t) = self.srgb_profile.create_transform_f64(
                Layout::Rgb,
                &cmyk_profile,
                Layout::Rgba,
                options,
            ) {
                self.reverse_cmyk_f64 = Some(t);
                return Some(());
            }
        }
        None
    }

    /// Pre-warm the lazy sRGB→CMYK reverse transform. Should be called once on
    /// the build thread that owns `&mut IccCache`, after the system CMYK
    /// profile has been registered, so that band renderers (which only hold
    /// `&IccCache`) can use [`Self::convert_rgb_to_cmyk_readonly`] without
    /// having to mutate state.
    pub fn prepare_reverse_cmyk(&mut self) {
        let _ = self.ensure_reverse_cmyk_transform();
    }

    /// Pre-build per-intent `Lab → OutputIntent CMYK` samplers from the
    /// document's OutputIntent profile. Call once after the OI is registered
    /// so [`Self::convert_lab_to_oi_cmyk`] can run from `&IccCache`.
    /// No-op when no default CMYK profile is registered, when the profile
    /// shape doesn't expose B2A LUTs (shaper-matrix CMYK with no LUT, etc.),
    /// or when proofing is disabled.
    pub fn prepare_lab_to_oi_cmyk(&mut self) {
        let Some(hash) = self.default_cmyk_hash else {
            return;
        };
        let Some(profile) = self.profiles.get(&hash).cloned() else {
            return;
        };
        use moxcms::RenderingIntent;
        for &intent in &[
            RenderingIntent::Perceptual,
            RenderingIntent::RelativeColorimetric,
            RenderingIntent::Saturation,
            RenderingIntent::AbsoluteColorimetric,
        ] {
            let i = intent as usize;
            if self.lab_to_oi_per_intent[i].is_some() {
                continue;
            }
            if let Some(sampler) = perceptual::LabToCmykSampler::build(&profile, intent) {
                self.lab_to_oi_per_intent[i] = Some(Arc::new(sampler));
            }
        }
    }

    /// Convert a PDF Lab triplet (L\* ∈ [0, 100], a\*/b\* ∈ [-128, 127]) to
    /// OutputIntent CMYK using the OI's per-intent B2A table. Returns `None`
    /// when the per-intent sampler hasn't been built (call
    /// [`Self::prepare_lab_to_oi_cmyk`] first) or when the intent slot is
    /// empty (falls back to Perceptual when available).
    ///
    /// Used by `lab_to_device_color` to populate `DeviceColor::native_cmyk`
    /// so the renderer's parallel CMYK buffer holds the same direct
    /// `Lab → OI CMYK` value Acrobat's ACE produces. Without this, the
    /// renderer falls back to `convert_rgb_to_cmyk_readonly` (sRGB → ICC
    /// reverse), which drifts visibly under CMYK-group blends. GWG 22.1's
    /// ColorBurn form over a Lab BG is the canonical surfacing case.
    pub fn convert_lab_to_oi_cmyk(
        &self,
        l_star: f64,
        a_star: f64,
        b_star: f64,
        intent: IccRenderingIntent,
    ) -> Option<[f64; 4]> {
        let sampler = self.lab_to_oi_per_intent[intent as usize]
            .as_ref()
            .or(self.lab_to_oi_per_intent[IccRenderingIntent::Perceptual as usize].as_ref())?;
        Some(sampler.sample_pdf_lab(l_star, a_star, b_star))
    }

    /// Convert an sRGB color to CMYK using the system CMYK profile, without
    /// mutating any state. Returns `None` when the reverse transform has not
    /// been pre-built (call [`Self::prepare_reverse_cmyk`] first) or when no
    /// system CMYK profile is registered.
    ///
    /// The returned components are clamped to `[0, 1]`.
    pub fn convert_rgb_to_cmyk_readonly(&self, r: f64, g: f64, b: f64) -> Option<[f64; 4]> {
        let reverse = self.reverse_cmyk_f64.as_ref()?;
        let src_rgb = [r.clamp(0.0, 1.0), g.clamp(0.0, 1.0), b.clamp(0.0, 1.0)];
        let mut cmyk = [0.0f64; 4];
        reverse.transform(&src_rgb, &mut cmyk).ok()?;
        Some([
            cmyk[0].clamp(0.0, 1.0),
            cmyk[1].clamp(0.0, 1.0),
            cmyk[2].clamp(0.0, 1.0),
            cmyk[3].clamp(0.0, 1.0),
        ])
    }

    /// Round-trip an RGB color through the system CMYK profile: sRGB→CMYK→sRGB.
    /// Used when compositing in a DeviceCMYK page group — saturated RGB colors
    /// become more muted after passing through the CMYK gamut.
    /// Returns None if no CMYK profile is loaded.
    pub fn round_trip_rgb_via_cmyk(&mut self, r: f64, g: f64, b: f64) -> Option<(f64, f64, f64)> {
        self.ensure_reverse_cmyk_transform()?;
        let hash = *self.default_cmyk_hash.as_ref()?;
        let reverse = self.reverse_cmyk_f64.as_ref()?;

        // sRGB → CMYK
        let src_rgb = [r.clamp(0.0, 1.0), g.clamp(0.0, 1.0), b.clamp(0.0, 1.0)];
        let mut cmyk = [0.0f64; 4];
        reverse.transform(&src_rgb, &mut cmyk).ok()?;

        // CMYK → sRGB (via existing forward transform)
        let forward = self.transforms.get(&hash)?;
        let mut dst = [0.0f64; 3];
        forward.transform_f64.transform(&cmyk, &mut dst).ok()?;

        let dst = bpc_post_correct(dst, forward.bpc_params.as_ref());
        Some((
            dst[0].clamp(0.0, 1.0),
            dst[1].clamp(0.0, 1.0),
            dst[2].clamp(0.0, 1.0),
        ))
    }

    /// Disable all ICC color management — clears all profiles, transforms,
    /// and caches. Equivalent to the CLI's `--no-icc` flag.
    pub fn disable(&mut self) {
        self.profiles.clear();
        self.transforms.clear();
        self.color_cache.clear();
        self.raw_bytes.clear();
        self.default_cmyk_hash = None;
        self.system_cmyk_bytes = None;
        self.reverse_cmyk_f64 = None;
    }
}

/// Bake a 4D CLUT by sampling an 8-bit CMYK→sRGB transform on a regular grid.
///
/// Generates `grid_n^4` CMYK sample points (each channel stepping `0..=255` in
/// `grid_n` steps), invokes moxcms once on the full batch, and stores the
/// packed sRGB output. Storage order is K outermost, then Y, M, C innermost,
/// matching the interpolation access pattern in `apply_clut4_cmyk_to_rgb`.
///
/// Returns `None` if the transform invocation fails — callers fall back to
/// direct moxcms calls per image.
fn bake_clut4(
    transform: &(dyn TransformExecutor<u8> + Send + Sync),
    grid_n: u8,
    bpc_params: Option<&BpcParams>,
) -> Option<Clut4> {
    let n = grid_n as usize;
    if !(2..=33).contains(&n) {
        return None;
    }
    let total = n * n * n * n;
    // Sample grid: for each (k, y, m, c) grid index, emit bytes (c, m, y, k).
    // moxcms consumes this as packed 4-channel input.
    let mut src = Vec::with_capacity(total * 4);
    let step = |i: usize| -> u8 {
        // Spread grid indices evenly across 0..=255 (endpoints inclusive).
        ((i as u32 * 255) / (n as u32 - 1)) as u8
    };
    for k in 0..n {
        let kv = step(k);
        for y in 0..n {
            let yv = step(y);
            for m in 0..n {
                let mv = step(m);
                for c in 0..n {
                    let cv = step(c);
                    src.extend_from_slice(&[cv, mv, yv, kv]);
                }
            }
        }
    }
    let mut dst = vec![0u8; total * 3];
    transform.transform(&src, &mut dst).ok()?;

    // Bake BPC into every grid point so runtime CLUT lookup stays at zero
    // per-pixel cost.
    if let Some(p) = bpc_params {
        for px in dst.chunks_exact_mut(3) {
            let out = apply_bpc_rgb_u8([px[0], px[1], px[2]], p);
            px[0] = out[0];
            px[1] = out[1];
            px[2] = out[2];
        }
    }

    Some(Clut4 {
        grid_n,
        data: Arc::new(dst),
    })
}

/// Single-pixel CMYK→sRGB lookup against the baked CLUT, with f64 inputs and
/// outputs. Performs the same K-bracket × 3D tetrahedral interpolation as the
/// bulk image path, but in floating point so a flat fill (which would otherwise
/// quantize the input to u8) matches a gradient stop's byte output to within
/// the CLUT's grid-interpolation error.
fn sample_clut4_single_f64(clut: &Clut4, c: f64, m: f64, y: f64, k: f64) -> (f64, f64, f64) {
    let n = clut.grid_n as usize;
    let nm1 = (n - 1) as f64;
    let lut = clut.data.as_slice();
    let stride_c: usize = 3;
    let stride_m: usize = n * stride_c;
    let stride_y: usize = n * stride_m;
    let stride_k: usize = n * stride_y;

    #[inline]
    fn axis(v: f64, nm1: f64, n: usize) -> (usize, usize, f64) {
        let scaled = v.clamp(0.0, 1.0) * nm1;
        let lo = scaled.floor();
        let frac = scaled - lo;
        let lo_i = lo as usize;
        let hi_i = (lo_i + 1).min(n - 1);
        (lo_i, hi_i, frac)
    }

    let (ci, ci1, fc) = axis(c, nm1, n);
    let (mi, mi1, fm) = axis(m, nm1, n);
    let (yi, yi1, fy) = axis(y, nm1, n);
    let (ki, ki1, fk) = axis(k, nm1, n);

    // Pick tetrahedron vertices (Kasson '94) — same logic as the u8 path.
    let (a_dxmy, b_dxmy, w1, w2, w3) = if fc >= fm {
        if fm >= fy {
            ((1, 0, 0), (1, 1, 0), fc, fm, fy)
        } else if fc >= fy {
            ((1, 0, 0), (1, 0, 1), fc, fy, fm)
        } else {
            ((0, 0, 1), (1, 0, 1), fy, fc, fm)
        }
    } else if fc >= fy {
        ((0, 1, 0), (1, 1, 0), fm, fc, fy)
    } else if fm >= fy {
        ((0, 1, 0), (0, 1, 1), fm, fy, fc)
    } else {
        ((0, 0, 1), (0, 1, 1), fy, fm, fc)
    };

    let corner = |d: (u8, u8, u8)| -> usize {
        let (dc, dm, dy) = d;
        let cx = if dc == 0 { ci } else { ci1 };
        let mx = if dm == 0 { mi } else { mi1 };
        let yx = if dy == 0 { yi } else { yi1 };
        yx * stride_y + mx * stride_m + cx * stride_c
    };

    let o000 = corner((0, 0, 0));
    let o111 = corner((1, 1, 1));
    let oa = corner(a_dxmy);
    let ob = corner(b_dxmy);

    let base_lo = ki * stride_k;
    let base_hi = ki1 * stride_k;

    let tetra_channel = |base: usize, ch: usize| -> f64 {
        let v000 = lut[base + o000 + ch] as f64;
        let va = lut[base + oa + ch] as f64;
        let vb = lut[base + ob + ch] as f64;
        let v111 = lut[base + o111 + ch] as f64;
        v000 + (va - v000) * w1 + (vb - va) * w2 + (v111 - vb) * w3
    };

    let r_lo = tetra_channel(base_lo, 0);
    let g_lo = tetra_channel(base_lo, 1);
    let b_lo = tetra_channel(base_lo, 2);
    let (r_hi, g_hi, b_hi) = if ki == ki1 {
        (r_lo, g_lo, b_lo)
    } else {
        (
            tetra_channel(base_hi, 0),
            tetra_channel(base_hi, 1),
            tetra_channel(base_hi, 2),
        )
    };

    let inv_fk = 1.0 - fk;
    let r = (r_lo * inv_fk + r_hi * fk) / 255.0;
    let g = (g_lo * inv_fk + g_hi * fk) / 255.0;
    let b = (b_lo * inv_fk + b_hi * fk) / 255.0;
    (r.clamp(0.0, 1.0), g.clamp(0.0, 1.0), b.clamp(0.0, 1.0))
}

/// Convert an 8-bit packed CMYK buffer to 8-bit packed sRGB using the baked
/// 4D CLUT. For each pixel: bracket the K axis into two slices, run 3D
/// tetrahedral (Kasson) interpolation on (C,M,Y) in each slice, then linearly
/// blend the two results by the K fraction.
///
/// This preserves the profile's behavior across the K axis (UCR/black-point
/// transitions) while giving image-rate throughput.
fn apply_clut4_cmyk_to_rgb(clut: &Clut4, src: &[u8], pixel_count: usize) -> Vec<u8> {
    let n = clut.grid_n as usize;
    let nm1 = (n - 1) as u32;
    let lut = clut.data.as_slice();

    // Strides in bytes within the flat LUT (K outermost, then Y, M; C innermost).
    let stride_c: usize = 3;
    let stride_m: usize = n * stride_c;
    let stride_y: usize = n * stride_m;
    let stride_k: usize = n * stride_y;

    let mut out = vec![0u8; pixel_count * 3];

    // Per-axis: quantize byte → (lo_idx, hi_idx, frac_in_0_255).
    #[inline(always)]
    fn axis(v: u8, nm1: u32) -> (usize, usize, u32) {
        let scaled = v as u32 * nm1;
        let lo = scaled / 255;
        let frac = scaled - lo * 255;
        let hi = if lo < nm1 { lo + 1 } else { lo };
        (lo as usize, hi as usize, frac)
    }

    for i in 0..pixel_count {
        let o = i * 4;
        let c = src[o];
        let m = src[o + 1];
        let y = src[o + 2];
        let k = src[o + 3];

        let (ci, ci1, fc) = axis(c, nm1);
        let (mi, mi1, fm) = axis(m, nm1);
        let (yi, yi1, fy) = axis(y, nm1);
        let (ki, ki1, fk) = axis(k, nm1);

        // Pick tetrahedron vertices and sorted weights ONCE per pixel
        // (previously done per channel — 3× waste). Kasson '94:
        //   out = V000 + (Va - V000)*w1 + (Vb - Va)*w2 + (V111 - Vb)*w3
        // with w1 >= w2 >= w3 and Va, Vb the two intermediate corners.
        let (a_dxmy, b_dxmy, w1, w2, w3) = if fc >= fm {
            if fm >= fy {
                // C,M,Y
                ((1, 0, 0), (1, 1, 0), fc, fm, fy)
            } else if fc >= fy {
                // C,Y,M
                ((1, 0, 0), (1, 0, 1), fc, fy, fm)
            } else {
                // Y,C,M
                ((0, 0, 1), (1, 0, 1), fy, fc, fm)
            }
        } else if fc >= fy {
            // M,C,Y
            ((0, 1, 0), (1, 1, 0), fm, fc, fy)
        } else if fm >= fy {
            // M,Y,C
            ((0, 1, 0), (0, 1, 1), fm, fy, fc)
        } else {
            // Y,M,C
            ((0, 0, 1), (0, 1, 1), fy, fm, fc)
        };

        // Map tetrahedron corner selector (dc, dm, dy) → LUT offset within a K slice.
        let corner = |d: (u8, u8, u8)| -> usize {
            let (dc, dm, dy) = d;
            let cx = if dc == 0 { ci } else { ci1 };
            let mx = if dm == 0 { mi } else { mi1 };
            let yx = if dy == 0 { yi } else { yi1 };
            yx * stride_y + mx * stride_m + cx * stride_c
        };

        let o000 = corner((0, 0, 0));
        let o111 = corner((1, 1, 1));
        let oa = corner(a_dxmy);
        let ob = corner(b_dxmy);

        // Two K slices, 3 channels. Compute inline (no closures, no per-channel branching).
        let base_lo = ki * stride_k;
        let base_hi = ki1 * stride_k;

        // Per-channel tetrahedral formula in integer:
        //   accum = v000*255 + (va - v000)*w1 + (vb - va)*w2 + (v111 - vb)*w3
        // accum is in units of (value * 255), in range [0, 255*255].
        let tetra_channel = |base: usize, ch: usize| -> i32 {
            let v000 = lut[base + o000 + ch] as i32;
            let va = lut[base + oa + ch] as i32;
            let vb = lut[base + ob + ch] as i32;
            let v111 = lut[base + o111 + ch] as i32;
            v000 * 255 + (va - v000) * w1 as i32 + (vb - va) * w2 as i32 + (v111 - vb) * w3 as i32
        };

        let r_lo = tetra_channel(base_lo, 0);
        let g_lo = tetra_channel(base_lo, 1);
        let b_lo = tetra_channel(base_lo, 2);
        let (r_hi, g_hi, b_hi) = if ki == ki1 {
            (r_lo, g_lo, b_lo)
        } else {
            (
                tetra_channel(base_hi, 0),
                tetra_channel(base_hi, 1),
                tetra_channel(base_hi, 2),
            )
        };

        // Linear blend across K slices and rescale to u8.
        let inv_fk = (255 - fk) as i32;
        let fk_i = fk as i32;
        let round = 255 * 255 / 2;
        let finish = |lo: i32, hi: i32| -> u8 {
            let combined = lo * inv_fk + hi * fk_i + round;
            let v = combined / (255 * 255);
            v.clamp(0, 255) as u8
        };

        let di = i * 3;
        out[di] = finish(r_lo, r_hi);
        out[di + 1] = finish(g_lo, g_hi);
        out[di + 2] = finish(b_lo, b_hi);
    }

    out
}

/// Validate a baked CLUT against the direct moxcms transform over a
/// pseudorandom sample of CMYK inputs. Reports median and max per-channel
/// deviation (in u8 units) to stderr. Invoked only when `STET_ICC_VERIFY` is
/// set in the environment.
fn verify_clut4(
    clut: &Clut4,
    transform: &(dyn TransformExecutor<u8> + Send + Sync),
    bpc_params: Option<&BpcParams>,
) {
    const N_SAMPLES: usize = 4096;
    let mut rng: u64 = 0xa8b3c4d5e6f70819;
    let mut next = || {
        rng = rng
            .wrapping_mul(6364136223846793005)
            .wrapping_add(1442695040888963407);
        rng
    };
    let mut cmyk = Vec::with_capacity(N_SAMPLES * 4);
    for _ in 0..N_SAMPLES {
        let r = next();
        cmyk.extend_from_slice(&[
            (r & 0xff) as u8,
            ((r >> 8) & 0xff) as u8,
            ((r >> 16) & 0xff) as u8,
            ((r >> 24) & 0xff) as u8,
        ]);
    }
    let mut reference = vec![0u8; N_SAMPLES * 3];
    if transform.transform(&cmyk, &mut reference).is_err() {
        eprintln!("[ICC VERIFY] reference transform failed");
        return;
    }
    // Mirror the CLUT bake's BPC step in the reference path so the
    // comparison measures interpolation error, not whether BPC was applied.
    if let Some(p) = bpc_params {
        for px in reference.chunks_exact_mut(3) {
            let out = apply_bpc_rgb_u8([px[0], px[1], px[2]], p);
            px[0] = out[0];
            px[1] = out[1];
            px[2] = out[2];
        }
    }
    let interp = apply_clut4_cmyk_to_rgb(clut, &cmyk, N_SAMPLES);
    // Per-pixel Euclidean distance in 8-bit sRGB (crude ΔE proxy).
    let mut dists: Vec<f64> = Vec::with_capacity(N_SAMPLES);
    let mut max_ch: u8 = 0;
    for i in 0..N_SAMPLES {
        let dr = interp[i * 3] as i32 - reference[i * 3] as i32;
        let dg = interp[i * 3 + 1] as i32 - reference[i * 3 + 1] as i32;
        let db = interp[i * 3 + 2] as i32 - reference[i * 3 + 2] as i32;
        let d = ((dr * dr + dg * dg + db * db) as f64).sqrt();
        dists.push(d);
        max_ch = max_ch
            .max(dr.unsigned_abs() as u8)
            .max(dg.unsigned_abs() as u8)
            .max(db.unsigned_abs() as u8);
    }
    dists.sort_by(|a, b| a.partial_cmp(b).unwrap());
    let median = dists[N_SAMPLES / 2];
    let p99 = dists[(N_SAMPLES * 99) / 100];
    let max = dists[N_SAMPLES - 1];
    eprintln!(
        "[ICC VERIFY] N=17 CLUT vs direct moxcms (sRGB u8): median={:.2}, p99={:.2}, max={:.2}, max_per_channel={}",
        median, p99, max, max_ch
    );
}

/// Search system paths for CMYK ICC profile bytes without parsing or logging.
///
/// Returns the raw bytes suitable for passing to the viewer for ICC-aware rendering.
pub fn find_system_cmyk_profile_bytes() -> Option<Arc<Vec<u8>>> {
    find_system_cmyk_profile().map(Arc::new)
}

/// Search common system paths for a CMYK ICC profile.
fn find_system_cmyk_profile() -> Option<Vec<u8>> {
    #[cfg(target_os = "linux")]
    {
        let paths = [
            "/usr/share/color/icc/ghostscript/default_cmyk.icc",
            "/usr/share/color/icc/ghostscript/ps_cmyk.icc",
            "/usr/share/color/icc/colord/FOGRA39L_coated.icc",
        ];
        for path in &paths {
            if let Ok(bytes) = std::fs::read(path) {
                return Some(bytes);
            }
        }
        // Glob for SWOP profiles
        if let Ok(entries) = glob::glob("/usr/share/color/icc/colord/SWOP*.icc") {
            for entry in entries.flatten() {
                if let Ok(bytes) = std::fs::read(&entry) {
                    return Some(bytes);
                }
            }
        }
    }

    #[cfg(target_os = "macos")]
    {
        let dirs = [
            "/Library/ColorSync/Profiles",
            "/System/Library/ColorSync/Profiles",
        ];
        if let Some(home) = std::env::var_os("HOME") {
            let home_dir = std::path::PathBuf::from(home).join("Library/ColorSync/Profiles");
            if let Some(bytes) = scan_dir_for_cmyk_icc(&home_dir) {
                return Some(bytes);
            }
        }
        for dir in &dirs {
            if let Some(bytes) = scan_dir_for_cmyk_icc(std::path::Path::new(dir)) {
                return Some(bytes);
            }
        }
    }

    #[cfg(target_os = "windows")]
    {
        if let Some(sysroot) = std::env::var_os("SYSTEMROOT") {
            let dir = std::path::PathBuf::from(sysroot).join("System32/spool/drivers/color");
            if let Some(bytes) = scan_dir_for_cmyk_icc(&dir) {
                return Some(bytes);
            }
        }
    }

    None
}

/// Scan a directory for ICC files with CMYK color space.
#[cfg(any(target_os = "macos", target_os = "windows"))]
fn scan_dir_for_cmyk_icc(dir: &std::path::Path) -> Option<Vec<u8>> {
    let entries = std::fs::read_dir(dir).ok()?;
    for entry in entries.flatten() {
        let path = entry.path();
        let ext = path.extension().and_then(|e| e.to_str()).unwrap_or("");
        if ext.eq_ignore_ascii_case("icc") || ext.eq_ignore_ascii_case("icm") {
            if let Ok(bytes) = std::fs::read(&path) {
                // Check ICC header: color space at offset 16, 'CMYK' = 0x434D594B
                if bytes.len() >= 20 && &bytes[16..20] == b"CMYK" {
                    return Some(bytes);
                }
            }
        }
    }
    None
}

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

    #[test]
    fn test_icc_cache_new() {
        let cache = IccCache::new();
        assert!(cache.default_cmyk_hash.is_none());
        assert!(cache.profiles.is_empty());
        assert_eq!(cache.bpc_mode(), BpcMode::Auto);
    }

    #[test]
    fn test_icc_cache_options_default_matches_new() {
        let cache = IccCache::new_with_options(IccCacheOptions::default());
        assert_eq!(cache.bpc_mode(), BpcMode::Auto);
        assert!(cache.default_cmyk_hash.is_none());
    }

    #[test]
    fn test_icc_cache_options_bpc_off() {
        let cache = IccCache::new_with_options(IccCacheOptions {
            bpc_mode: BpcMode::Off,
            source_cmyk_profile: None,
        });
        assert_eq!(cache.bpc_mode(), BpcMode::Off);
        assert!(!cache.bpc_mode().is_enabled());
    }

    #[test]
    fn test_icc_cache_options_preloads_cmyk_profile() {
        // Skip when no system CMYK profile is available.
        let Some(cmyk_bytes) = find_system_cmyk_profile() else {
            return;
        };
        let cache = IccCache::new_with_options(IccCacheOptions {
            bpc_mode: BpcMode::On,
            source_cmyk_profile: Some(cmyk_bytes.clone()),
        });
        assert!(cache.default_cmyk_hash().is_some());
        assert_eq!(cache.bpc_mode(), BpcMode::On);
    }

    #[test]
    fn test_bpc_darkens_pure_k_per_color() {
        // Skip when no system CMYK profile is available.
        let Some(cmyk_bytes) = find_system_cmyk_profile() else {
            return;
        };

        // Without BPC: K=1 through default_cmyk.icc lands near RGB(55, 53, 53)
        // — the profile's as-mapped black projected through moxcms's sRGB B2A.
        let mut off = IccCache::new_with_options(IccCacheOptions {
            bpc_mode: BpcMode::Off,
            source_cmyk_profile: Some(cmyk_bytes.clone()),
        });
        let off_rgb = off.convert_cmyk(0.0, 0.0, 0.0, 1.0).unwrap();

        // With BPC: K=1 must land significantly darker. Adobe Acrobat (lcms2)
        // produces RGB(35, 31, 32). moxcms's sRGB B2A handles very-dark XYZ
        // slightly differently from lcms2, so our post-correct lands a few
        // levels brighter than Acrobat — the meaningful invariant is "K=1 is
        // visibly darker than the no-BPC baseline by a substantial margin."
        let mut on = IccCache::new_with_options(IccCacheOptions {
            bpc_mode: BpcMode::On,
            source_cmyk_profile: Some(cmyk_bytes),
        });
        let on_rgb = on.convert_cmyk(0.0, 0.0, 0.0, 1.0).unwrap();

        // Precondition: this test only exercises BPC's darkening effect, which
        // requires a profile whose black point has non-zero luminance. Some
        // system-supplied CMYK profiles (e.g. macOS's default ColorSync CMYK)
        // already map K=1 to (near-)zero XYZ, so BPC has nothing to correct
        // and off == on. Skip in that case — there's no regression to anchor
        // here, just a profile that doesn't benefit from BPC.
        if (on_rgb.1 - off_rgb.1).abs() < 0.005 {
            eprintln!(
                "Skipping: system CMYK profile's black point is already ~zero; \
                 BPC is a no-op here. off={off_rgb:?} on={on_rgb:?}"
            );
            return;
        }

        assert!(
            on_rgb.1 + 0.03 < off_rgb.1,
            "BPC should darken K=1 by ≥0.03 (~8 RGB levels): off={off_rgb:?} on={on_rgb:?}"
        );
        // And the resulting RGB should be in the "deep gray" range — well
        // under 0.25 (RGB ≤ ~64) on every channel.
        assert!(
            on_rgb.0 < 0.25 && on_rgb.1 < 0.25 && on_rgb.2 < 0.25,
            "Expected deep gray after BPC, got {on_rgb:?}"
        );
    }

    #[test]
    fn test_bpc_white_anchored_per_color() {
        // Skip when no system CMYK profile is available.
        let Some(cmyk_bytes) = find_system_cmyk_profile() else {
            return;
        };
        let mut cache = IccCache::new_with_options(IccCacheOptions {
            bpc_mode: BpcMode::On,
            source_cmyk_profile: Some(cmyk_bytes),
        });
        // CMYK white (no ink) must still render as sRGB white under BPC.
        let (r, g, b) = cache.convert_cmyk(0.0, 0.0, 0.0, 0.0).unwrap();
        assert!(r > 0.99 && g > 0.99 && b > 0.99, "({r}, {g}, {b})");
    }

    #[test]
    fn test_bpc_image_clut_path_darkens_pure_k() {
        // Skip when no system CMYK profile is available.
        let Some(cmyk_bytes) = find_system_cmyk_profile() else {
            return;
        };

        // Build a 1-pixel CMYK image at K=1 and route it through the CLUT.
        // Without BPC vs with BPC, the K=1 pixel must shift darker, mirroring
        // the per-color path behaviour.
        let off = IccCache::new_with_options(IccCacheOptions {
            bpc_mode: BpcMode::Off,
            source_cmyk_profile: Some(cmyk_bytes.clone()),
        });
        let on = IccCache::new_with_options(IccCacheOptions {
            bpc_mode: BpcMode::On,
            source_cmyk_profile: Some(cmyk_bytes),
        });
        let off_hash = *off.default_cmyk_hash().unwrap();
        let on_hash = *on.default_cmyk_hash().unwrap();

        let pixel = [0u8, 0, 0, 255]; // C=0 M=0 Y=0 K=255
        let off_rgb = off.convert_image_8bit(&off_hash, &pixel, 1).unwrap();
        let on_rgb = on.convert_image_8bit(&on_hash, &pixel, 1).unwrap();

        // Precondition, same as `test_bpc_darkens_pure_k_per_color`: BPC only
        // shifts pixels when the profile's black point has non-zero luminance.
        // Skip when the system profile already maps K=1 to near-zero XYZ.
        if (on_rgb[1] as i32 - off_rgb[1] as i32).abs() <= 1 {
            eprintln!(
                "Skipping: system CMYK profile's black point is already ~zero; \
                 BPC is a no-op here. off={off_rgb:?} on={on_rgb:?}"
            );
            return;
        }

        // BPC must darken the green channel by ≥8 RGB levels (mirrors the
        // per-color path's anchor in test_bpc_darkens_pure_k_per_color).
        assert!(
            (on_rgb[1] as i32) + 8 < (off_rgb[1] as i32),
            "CLUT BPC should darken K=1 image green by ≥8 levels: off={off_rgb:?} on={on_rgb:?}"
        );
        // And land in the deep-gray range.
        assert!(
            on_rgb[0] < 64 && on_rgb[1] < 64 && on_rgb[2] < 64,
            "Expected deep gray after CLUT BPC, got {on_rgb:?}"
        );
    }

    #[test]
    fn test_bpc_off_image_matches_per_color_off() {
        // With --bpc off, the bulk image path's K=1 output must match the
        // per-color path's K=1 output (within u8 quantization). Anchors that
        // disabling BPC reproduces stet's pre-fix behaviour bit-for-bit on
        // the dominant CMYK image path.
        let Some(cmyk_bytes) = find_system_cmyk_profile() else {
            return;
        };
        let mut cache = IccCache::new_with_options(IccCacheOptions {
            bpc_mode: BpcMode::Off,
            source_cmyk_profile: Some(cmyk_bytes),
        });
        let hash = *cache.default_cmyk_hash().unwrap();

        let pixel = [0u8, 0, 0, 255];
        let img = cache.convert_image_8bit(&hash, &pixel, 1).unwrap();
        let (r, g, b) = cache.convert_cmyk(0.0, 0.0, 0.0, 1.0).unwrap();
        let pc = [
            (r * 255.0).round() as i32,
            (g * 255.0).round() as i32,
            (b * 255.0).round() as i32,
        ];
        // CLUT interpolation drift can introduce ±1 vs the direct f64 path.
        assert!((img[0] as i32 - pc[0]).abs() <= 2, "img={img:?} pc={pc:?}");
        assert!((img[1] as i32 - pc[1]).abs() <= 2, "img={img:?} pc={pc:?}");
        assert!((img[2] as i32 - pc[2]).abs() <= 2, "img={img:?} pc={pc:?}");
    }

    #[test]
    fn test_register_invalid_profile() {
        let mut cache = IccCache::new();
        assert!(cache.register_profile(b"not a valid ICC profile").is_none());
    }

    #[test]
    fn test_srgb_identity_transform() {
        // Create an sRGB profile, register it, and verify identity-ish conversion
        let srgb = ColorProfile::new_srgb();
        let bytes = srgb.encode().unwrap();
        let mut cache = IccCache::new();
        let hash = cache.register_profile(&bytes).unwrap();

        // Red should stay approximately red
        let (r, g, b) = cache.convert_color(&hash, &[1.0, 0.0, 0.0]).unwrap();
        assert!((r - 1.0).abs() < 0.02, "r={r}");
        assert!(g < 0.02, "g={g}");
        assert!(b < 0.02, "b={b}");

        // White
        let (r, g, b) = cache.convert_color(&hash, &[1.0, 1.0, 1.0]).unwrap();
        assert!((r - 1.0).abs() < 0.02);
        assert!((g - 1.0).abs() < 0.02);
        assert!((b - 1.0).abs() < 0.02);
    }

    #[test]
    fn test_srgb_image_transform() {
        let srgb = ColorProfile::new_srgb();
        let bytes = srgb.encode().unwrap();
        let mut cache = IccCache::new();
        let hash = cache.register_profile(&bytes).unwrap();

        // 2 pixels: red, green
        let src = [255u8, 0, 0, 0, 255, 0];
        let result = cache.convert_image_8bit(&hash, &src, 2).unwrap();
        assert_eq!(result.len(), 6);
        // Red pixel should be approximately (255, 0, 0)
        assert!(result[0] > 240);
        assert!(result[1] < 15);
        assert!(result[2] < 15);
    }

    #[test]
    fn test_convert_rgb_to_cmyk_readonly() {
        // Skip when no system CMYK profile is available (CI without ICC packs).
        let Some(cmyk_bytes) = find_system_cmyk_profile() else {
            return;
        };
        let mut cache = IccCache::new();
        let hash = cache.register_profile(&cmyk_bytes).unwrap();
        cache.set_default_cmyk_hash(hash);

        // Before pre-warming the reverse transform must be unavailable.
        assert!(cache.convert_rgb_to_cmyk_readonly(0.0, 0.0, 0.0).is_none());

        cache.prepare_reverse_cmyk();

        // Pure black sRGB should land deep in K (any reasonable CMYK profile
        // produces a high K component).
        let cmyk = cache
            .convert_rgb_to_cmyk_readonly(0.0, 0.0, 0.0)
            .expect("reverse transform should be available after prepare");
        assert!(
            cmyk[3] > 0.5,
            "expected K>0.5 for sRGB black, got cmyk={cmyk:?}"
        );

        // Pure white sRGB should land near (0,0,0,0) — minimal ink.
        let cmyk = cache
            .convert_rgb_to_cmyk_readonly(1.0, 1.0, 1.0)
            .expect("reverse transform should be available");
        for (i, v) in cmyk.iter().enumerate() {
            assert!(*v < 0.05, "expected near-zero ink at chan {i}, got {v}");
        }
    }

    /// White CMYK (0,0,0,0) routed through the perceptual CLUT must land at
    /// pure white sRGB. Catches scaling errors in the PCS-Lab decode (the
    /// most likely place to introduce a uniform brightness shift).
    #[test]
    fn test_perceptual_clut_white_anchor() {
        let Some(cmyk_bytes) = find_system_cmyk_profile() else {
            return;
        };
        let mut cache = IccCache::new_with_options(IccCacheOptions {
            bpc_mode: BpcMode::Off,
            source_cmyk_profile: Some(cmyk_bytes),
        });
        let rgb = cache.convert_cmyk(0.0, 0.0, 0.0, 0.0).unwrap();
        // Tolerate a few u8 levels of grid-interpolation drift; the bound
        // catches order-of-magnitude bugs (a ~0.5x scale would land at ~127).
        assert!(
            rgb.0 > 0.97 && rgb.1 > 0.97 && rgb.2 > 0.97,
            "CMYK white should map near sRGB white, got {rgb:?}"
        );
    }

    /// Single-color CLUT lookup must agree with bulk image conversion to
    /// within u8 quantization on the same input. Anchors the requirement
    /// that flat CMYK fills match adjacent gradient stops byte-for-byte.
    #[test]
    fn test_perceptual_clut_single_matches_bulk() {
        let Some(cmyk_bytes) = find_system_cmyk_profile() else {
            return;
        };
        let mut cache = IccCache::new_with_options(IccCacheOptions {
            bpc_mode: BpcMode::On,
            source_cmyk_profile: Some(cmyk_bytes),
        });
        let hash = *cache.default_cmyk_hash().unwrap();
        // A handful of saturated and midtone CMYK samples.
        let samples: &[(u8, u8, u8, u8)] = &[
            (0, 0, 0, 0),
            (255, 0, 0, 0),
            (0, 255, 0, 0),
            (0, 0, 255, 0),
            (0, 0, 0, 255),
            (38, 255, 255, 0), // ≈ (0.15, 1.0, 1.0, 0.0) — the GWG demo input
            (128, 128, 128, 0),
        ];
        let pixels: Vec<u8> = samples
            .iter()
            .flat_map(|&(c, m, y, k)| [c, m, y, k])
            .collect();
        let bulk = cache
            .convert_image_8bit(&hash, &pixels, samples.len())
            .unwrap();
        for (i, &(c, m, y, k)) in samples.iter().enumerate() {
            let single = cache
                .convert_color(
                    &hash,
                    &[
                        c as f64 / 255.0,
                        m as f64 / 255.0,
                        y as f64 / 255.0,
                        k as f64 / 255.0,
                    ],
                )
                .unwrap();
            let single_u8 = [
                (single.0 * 255.0).round() as i32,
                (single.1 * 255.0).round() as i32,
                (single.2 * 255.0).round() as i32,
            ];
            let b = &bulk[i * 3..i * 3 + 3];
            for ch in 0..3 {
                let delta = (single_u8[ch] - b[ch] as i32).abs();
                assert!(
                    delta <= 2,
                    "single vs bulk mismatch at sample {i} chan {ch}: \
                     single={single_u8:?} bulk={:?} delta={delta}",
                    [b[0], b[1], b[2]]
                );
            }
        }
    }
}