rlx_core/render/scenes/

reaction_diffusion.rs

//! Reaction-diffusion scene: a Gray-Scott simulation evolving on a fixed
//! internal grid via the reusable [`PingPongField`](crate::render::feedback)
//! (ADR-0012). The engine's first *stateful* scene — each frame's field depends
//! on the previous frame's, held in a texture and stepped by a simulation
//! shader — unlocking the organic, restructuring look (nested contours, cellular
//! tissue, a hatched maze) stateless scenes can't produce.
//!
//! Phase 1 is a walking skeleton: a fixed number of sub-steps per frame and a
//! grayscale present. The fixed-timestep accumulator (Phase 2), audio-reactive
//! named parameters (Phase 3), and the iso-contour/hatch look (Phase 4) land on
//! top of this. All randomness is the seeded initial scatter (NFR 6): the field
//! is a pure function of the seed + the fixed-`dt` step sequence.
//!
//! **GPU resources are built lazily, on first render.** The scene stores a
//! device handle and constructs its pipelines/textures only when it is first
//! drawn (see `Resources`). This keeps the resources off the device until the
//! scene is actually shown, and — importantly — lets the headless capture tests
//! build the full roster on the software adapter: WARP mis-renders the
//! pre-existing fragment-field pipeline once this scene's *full* set of feedback
//! resources coexists on the device (a cumulative software-rasterizer quirk with
//! no wgpu validation error; real hardware is unaffected; a WARP reading,
//! unverified on lavapipe as of 2026-10-06). Deferring
//! construction means a capture that never activates this scene never builds
//! those resources, so the other scenes' captures stay correct; a capture that
//! *does* activate it builds them and renders this scene normally.

// Hot-path panic-denial pragma (Plan 0002 Phase 2, extended to scenes by Plan
// 0003 Phase 0). Encodes its passes every displayed frame.
#![deny(
    clippy::unwrap_used,
    clippy::expect_used,
    clippy::indexing_slicing,
    clippy::panic,
    clippy::unreachable
)]

use crate::render::gpu;

use super::common;
use super::{Scene, SeededRng};
use crate::dsp::AnalysisFrame;
use crate::render::feedback::PingPongField;
use crate::render::palette::{self, Palette};
use crate::render::scenes::{ParamGroup, ParamKind, ParamSpec, default_of};

/// Fixed internal simulation grid (square). 256² resolves the Gray-Scott
/// patterns well while staying cheap enough that the headless capture tests run
/// briskly on a software adapter — a 512² grid quadruples the per-step
/// fragment work the differential tests pay each warm-up frame (ADR-0012 gives
/// 512² only as an example).
const GRID: u32 = 256;

/// Wall-clock duration of one Gray-Scott sub-step (Plan 0014 Phase 2). The
/// fixed-timestep accumulator runs one sub-step per `FIXED_STEP` of injected
/// real `dt`, so the simulation evolves at the same rate on any refresh — at the
/// live/​capture `dt` of 1/60 s this is 12 sub-steps per frame.
const FIXED_STEP: f32 = 1.0 / 720.0;

/// Max sub-steps encoded in a single frame. A long stall would otherwise queue
/// unbounded work (accumulator spiral-of-death); past this the accumulator's
/// backlog is dropped, so the sim briefly slows rather than diverging (ADR-0012).
/// 40 covers a ~55 ms hitch before it bites.
const MAX_SUBSTEPS: u32 = 40;

/// Seeded initial-scatter blobs, and the uniform array's capacity.
const SEED_BLOBS: usize = 30;
const MAX_BLOBS: usize = 32;
// The ASCII bytes of "LMV_RD_1" read as a number. Re-spelling them to
// match a renamed prefix changes the seeded scatter and moves this
// scene's goldens, so the value is opaque and stays as it is.
const SEED: u64 = 0x4C4D_565F_5244_5F31;

/// Parameter defaults — the "mitosis" Gray-Scott regime (Pearson's
/// classification): spots that perpetually divide, so the field keeps
/// restructuring rather than settling into a static pattern.
const DEFAULT_FEED: f32 = default_of(PARAMS, "feed");
const DEFAULT_KILL: f32 = default_of(PARAMS, "kill");
/// Diffusion rates for the two species (classic Karl Sims values at internal
/// `dt = 1`, paired with the 3×3 Laplacian kernel in the shader). The `flow`
/// param (Phase 3) scales both, keeping their ratio, so a band can coarsen or
/// tighten the pattern's spatial scale.
const DIFFUSE_U: f32 = 0.16;
const DIFFUSE_V: f32 = 0.08;
/// `flow` default: unscaled diffusion.
const DEFAULT_FLOW: f32 = default_of(PARAMS, "flow");

/// Present-look defaults (Phase 4): palette hue offset, iso-contour band count,
/// hatch stripe spacing in texels, and glow strength.
const DEFAULT_HUE: f32 = 0.0;
const DEFAULT_CONTOUR: f32 = default_of(PARAMS, "contour");
const DEFAULT_HATCH: f32 = default_of(PARAMS, "hatch");
const DEFAULT_GLOW: f32 = default_of(PARAMS, "glow");
// Shared palette color knobs (ADR-0021 / Plan 0020 Phase 5). `color_span` = 0.85
// (the old fixed field-to-gradient coefficient) + `color_center` = 0 +
// `saturation` = 1 + `palette_mix` = 0 reproduce the prior present-look color math
// (now sampling the shared LUT instead of the private cosine).
const DEFAULT_COLOR_SPAN: f32 = default_of(PARAMS, "color_span");
const DEFAULT_COLOR_CENTER: f32 = default_of(PARAMS, "color_center");
/// View transform defaults (ADR-0018): identity — `zoom` = 1 leaves the sampled
/// window unscaled, `pan` = 0 unshifted, so an unbound preset is byte-unchanged.
const DEFAULT_ZOOM: f32 = 1.0;

/// Beat-stamped seed injection (Phase 3). A rising `inject` edge stamps a blob
/// of V into the field at the next seeded position, so a beat spawns new growth.
/// Positions come from a `SeededRng` (NFR 6) so a capture reproduces exactly.
const INJECT_RADIUS: f32 = 0.045;
const INJECT_AMOUNT: f32 = 0.85;
const INJECT_SEED: u64 = 0x4C4D_5244_494E_4A31; // "LMRDINJ1"
/// `inject` rises past this to fire one stamp (edge-triggered, not per-frame).
const INJECT_THRESHOLD: f32 = 0.5;

/// Seed pass: U = 1 everywhere, V = 1 inside the scattered blobs.
const INIT_SHADER: &str = r#"
struct Init {
    blobs: array<vec4<f32>, 32>, // xy: center (uv), z: radius, w: unused
    count: vec4<u32>,            // x: active blob count
}
@group(0) @binding(0) var<uniform> init: Init;

@fragment
fn fs_main(in: VsOut) -> @location(0) vec4<f32> {
    var v = 0.0;
    let n = init.count.x;
    for (var i = 0u; i < n; i = i + 1u) {
        let b = init.blobs[i];
        if (distance(in.uv, b.xy) < b.z) {
            v = 1.0;
        }
    }
    return vec4<f32>(1.0 - v, v, 0.0, 1.0);
}
"#;

/// Sim pass: one Gray-Scott step, reading the previous field, writing the next.
const SIM_SHADER: &str = r#"
struct Sim {
    p: vec4<f32>,   // x: feed, y: kill, z: diffuse_u, w: diffuse_v
    inj: vec4<f32>, // xy: stamp center (uv), z: radius, w: amount (0 = no stamp)
}
@group(0) @binding(0) var<uniform> sim: Sim;
@group(0) @binding(1) var field: texture_2d<f32>;

// Toroidal texel fetch (wrap at the edges) of the (U, V) pair.
fn ld(c: vec2<i32>, size: vec2<i32>) -> vec2<f32> {
    let w = ((c % size) + size) % size;
    return textureLoad(field, w, 0).xy;
}

@fragment
fn fs_main(in: VsOut) -> @location(0) vec4<f32> {
    let size = vec2<i32>(textureDimensions(field));
    let c = vec2<i32>(i32(in.pos.x), i32(in.pos.y));
    let m = ld(c, size);
    let u = m.x;
    let v = m.y;

    // 3×3 Laplacian: orthogonal 0.2, diagonal 0.05, center -1.
    var lap = ld(c + vec2<i32>(-1, 0), size) * 0.2
        + ld(c + vec2<i32>(1, 0), size) * 0.2
        + ld(c + vec2<i32>(0, -1), size) * 0.2
        + ld(c + vec2<i32>(0, 1), size) * 0.2
        + ld(c + vec2<i32>(-1, -1), size) * 0.05
        + ld(c + vec2<i32>(1, -1), size) * 0.05
        + ld(c + vec2<i32>(-1, 1), size) * 0.05
        + ld(c + vec2<i32>(1, 1), size) * 0.05;
    lap = lap - m;

    let feed = sim.p.x;
    let kill = sim.p.y;
    let du = sim.p.z;
    let dv = sim.p.w;
    let reaction = u * v * v;
    let nu = u + du * lap.x - reaction + feed * (1.0 - u);
    var nv = v + dv * lap.y + reaction - (kill + feed) * v;

    // Beat-stamped seed injection (Phase 3), folded into the sim so no extra
    // pipeline is needed. `inj.w` is non-zero only on the stamp frame; it is
    // applied on every sub-step of that frame, so V saturates at the stamp.
    let stamp = sim.inj.w * (1.0 - smoothstep(sim.inj.z * 0.4, sim.inj.z, distance(in.uv, sim.inj.xy)));
    nv = nv + stamp;

    return vec4<f32>(clamp(nu, 0.0, 1.0), clamp(nv, 0.0, 1.0), 0.0, 1.0);
}
"#;

/// Present pass (Phase 4): the reference aesthetic — analytic iso-contours of
/// the V field with `fwidth` anti-aliasing, a cosine palette coloring the nested
/// loops, gradient-aligned hatch/comb ticks, and a soft glow.
const PRESENT_SHADER: &str = r#"
struct Present {
    // x: hue, y: contour density, z: hatch frequency (texels), w: glow
    a: vec4<f32>,
    // x: color_span, y: color_center, z: saturation, w: palette_mix
    b: vec4<f32>,
    // x: zoom, yz: pan (field-space view transform, ADR-0018), w: occlude (ADR-0085)
    c: vec4<f32>,
    // x: palette_steps (integral, quantized CPU-side), y: palette_contour
    // (ADR-0078), z: palette_contour_style (integral, rounded CPU-side),
    // w: palette_contour_ink (ADR-0197)
    d: vec4<f32>,
}
@group(0) @binding(0) var present_field: texture_2d<f32>;
@group(0) @binding(1) var present_samp: sampler;
@group(0) @binding(2) var<uniform> pp: Present;
// Shared gradient LUTs (ADR-0021): A/B for the `palette_mix` crossfade, one
// repeat sampler. Kept in this present bind group (a unique 6-entry layout) so it
// never matches another pipeline's layout on the DX12 WARP software adapter
// (where matching layouts alias; a WARP reading, unverified on lavapipe as of
// 2026-10-06).
@group(0) @binding(3) var lut_a: texture_2d<f32>;
@group(0) @binding(4) var lut_b: texture_2d<f32>;
@group(0) @binding(5) var lut_samp: sampler;

// Shared `saturation` (mirrors core/src/render/palette.rs::desaturate verbatim).
fn apply_saturation(c: vec3<f32>, s: f32) -> vec3<f32> {
    let luma = dot(c, vec3<f32>(0.299, 0.587, 0.114));
    return vec3<f32>(luma) + (c - vec3<f32>(luma)) * s;
}

// Shared `palette_steps` (mirrors core/src/render/palette.rs::band_coord
// verbatim, ADR-0078): snap the palette coordinate to a band centre before the
// LUT read. Below 1.5 steps it is the exact identity, not a one-band degenerate.
fn band_coord(t: f32, steps: f32) -> f32 {
    if (steps < 1.5) {
        return t;
    }
    return (floor(t * steps) + 0.5) / steps;
}

// Shared `palette_contour` (ADR-0078 / ADR-0133; the WGSL is the implementation,
// copied verbatim at each fragment-stage site — palette.rs has no CPU
// counterpart to be canonical, since `fwidth` exists only here).
//
// Darkens within one PIXEL of a band edge, so the line has the same weight where
// the field is shallow and where it is steep — AND ONLY WHERE THE INK ACTUALLY
// CHANGES (ADR-0133). It samples the two band centres either side of the nearest
// edge and returns unchanged when they resolve to the same colour within half a
// code value, which is below the LUT's own 8-bit quantization. On a smooth
// palette two distinct centres always differ by at least one code value, so
// every edge draws exactly as it did at any `palette_steps`; inside a plateau
// the LUT is literally constant and the samples are bit-equal, so the line
// vanishes there and survives at the run boundaries. One rule, both behaviours,
// no new parameter.
//
// The two LUTs, the sampler and `palette_mix` are EXPLICIT parameters rather
// than module-scope globals this happens to find: all six sites name them the
// same today, so implicit capture would compile — and would silently bind the
// shared function to whatever a future site called its textures.
//
// `textureSampleLevel`, not `textureSample`: the LUT has one mip, and an
// explicit LOD keeps these reads free of the uniformity requirement that a
// sample after a conditional return would otherwise carry.
//
// **What the line is drawn in is `style`** (ADR-0197): `0` the soft darkening
// above, `1` a hard darkening over the same footprint, `2` a soft line in the
// palette's own colour at `ink_t` and `3` a hard one. `style` arrives rounded to
// a whole number from the CPU (`palette::band_contour_style`), so the equality
// comparisons below are exact. The `style < 0.5` arm is the expression that
// shipped before the other three existed, which is what keeps every golden still.
fn band_contour_ink(
    col: vec3<f32>,
    t: f32,
    steps: f32,
    amount: f32,
    style: f32,
    ink_t: f32,
    lut_a: texture_2d<f32>,
    lut_b: texture_2d<f32>,
    lut_samp: sampler,
    mix_ab: f32,
) -> vec3<f32> {
    let f = t * steps;
    let w = max(fwidth(f), 1e-5);
    if (steps < 1.5 || amount <= 0.0) {
        return col;
    }
    let n = round(f);
    let m = clamp(mix_ab, 0.0, 1.0);
    let lo = mix(
        textureSampleLevel(lut_a, lut_samp, vec2<f32>((n - 0.5) / steps, 0.5), 0.0).rgb,
        textureSampleLevel(lut_b, lut_samp, vec2<f32>((n - 0.5) / steps, 0.5), 0.0).rgb,
        m
    );
    let hi = mix(
        textureSampleLevel(lut_a, lut_samp, vec2<f32>((n + 0.5) / steps, 0.5), 0.0).rgb,
        textureSampleLevel(lut_b, lut_samp, vec2<f32>((n + 0.5) / steps, 0.5), 0.0).rgb,
        m
    );
    if (all(abs(hi - lo) < vec3<f32>(0.5 / 255.0))) {
        return col;
    }
    let d = min(fract(f), 1.0 - fract(f));
    if (style < 0.5) {
        return col * (1.0 - clamp(amount, 0.0, 1.0) * (1.0 - smoothstep(0.0, w, d)));
    }
    let hard = style == 1.0 || style == 3.0;
    let cover = select(1.0 - smoothstep(0.0, w, d), f32(d < w), hard);
    let ink_lut = mix(
        textureSampleLevel(lut_a, lut_samp, vec2<f32>(ink_t, 0.5), 0.0).rgb,
        textureSampleLevel(lut_b, lut_samp, vec2<f32>(ink_t, 0.5), 0.0).rgb,
        m
    );
    let ink = select(vec3<f32>(0.0), ink_lut, style >= 2.0);
    return mix(col, ink, clamp(amount, 0.0, 1.0) * cover);
}

fn tap_v(uv: vec2<f32>) -> f32 {
    return textureSampleLevel(present_field, present_samp, uv, 0.0).y;
}

// C1 reconstruction of the finite field (Plan 0033 Phase 3, ADR-0034).
//
// Hardware bilinear is C0: the value is continuous, its gradient is not. This
// pass runs analytic iso-contours and a central-difference gradient over exactly
// that field, and `line_d` divides by `fwidth`, so a slope discontinuity far
// below the 8-bit output quantum is amplified into a visible tangent kink —
// which is why an 8x upscale of a smooth field read as angular facets.
//
// Catmull-Rom, not a smoothed texel coordinate. Warping the fractional
// coordinate by a quintic is the cheap trick and it is *wrong here*: a
// smoothstep-family warp has zero derivative at both ends, so it pins the
// reconstruction's gradient to zero at every texel centre. That is C1, but the
// derivative then oscillates once per cell, and a pass with this much gradient
// gain renders it as one scalloped step per texel — measurably worse than the
// faceting it replaces. Only a genuine higher-order filter has a smooth,
// non-degenerate derivative.
//
// Nine taps rather than sixteen: each pair of neighbouring weights is folded
// into one hardware-bilinear fetch at the weighted midpoint (`offset12`), which
// is exact for a separable cubic. `w1 + w2 >= 1` over the whole cell, so the
// division is safe.
//
// **Every** read of the field goes through here — value, gradient, contour and
// hatch — because fixing only the value tap fixes nothing: the gradient is where
// the discontinuity becomes visible.
fn sample_v(uv: vec2<f32>) -> f32 {
    let dims = vec2<f32>(textureDimensions(present_field));
    let sample_pos = uv * dims;
    let pos1 = floor(sample_pos - 0.5) + 0.5;
    let f = sample_pos - pos1;

    let w0 = f * (-0.5 + f * (1.0 - 0.5 * f));
    let w1 = 1.0 + f * f * (-2.5 + 1.5 * f);
    let w2 = f * (0.5 + f * (2.0 - 1.5 * f));
    let w3 = f * f * (-0.5 + 0.5 * f);
    let w12 = w1 + w2;

    let p0 = (pos1 - 1.0) / dims;
    let p3 = (pos1 + 2.0) / dims;
    let p12 = (pos1 + w2 / w12) / dims;

    var acc = 0.0;
    acc = acc + tap_v(vec2<f32>(p0.x, p0.y)) * w0.x * w0.y;
    acc = acc + tap_v(vec2<f32>(p12.x, p0.y)) * w12.x * w0.y;
    acc = acc + tap_v(vec2<f32>(p3.x, p0.y)) * w3.x * w0.y;

    acc = acc + tap_v(vec2<f32>(p0.x, p12.y)) * w0.x * w12.y;
    acc = acc + tap_v(vec2<f32>(p12.x, p12.y)) * w12.x * w12.y;
    acc = acc + tap_v(vec2<f32>(p3.x, p12.y)) * w3.x * w12.y;

    acc = acc + tap_v(vec2<f32>(p0.x, p3.y)) * w0.x * w3.y;
    acc = acc + tap_v(vec2<f32>(p12.x, p3.y)) * w12.x * w3.y;
    acc = acc + tap_v(vec2<f32>(p3.x, p3.y)) * w3.x * w3.y;
    return acc;
}

@fragment
fn fs_main(in: VsOut) -> @location(0) vec4<f32> {
    let dims = vec2<f32>(textureDimensions(present_field));
    let texel = 1.0 / dims;

    // View transform (ADR-0018): scale the sampled window about its centre by
    // `zoom`, then offset by `pan`. Default zoom = 1, pan = 0 leaves `in.uv`
    // untouched, so an unbound preset renders identically. Same shape fragment_field
    // uses for its field-space zoom/pan.
    let zoom = pp.c.x;
    // `pan.y` is negated because `in.uv` is now Y-flipped (ADR-0070): this pass
    // moved from the retired unflipped prelude, which reversed the direction a
    // positive `pan_y` scrolls the field. Every other scene applies pan in clip
    // space, where +y is up, and all four RD presets were authored against that
    // agreement — so the sign restores the shipped behaviour rather than changing
    // it. Measured both ways: pan_y = +0.12 moves the field 86 px, up before this
    // and down without this negation.
    let pan = vec2<f32>(pp.c.y, -pp.c.z);
    let uv = (in.uv - vec2<f32>(0.5, 0.5)) * zoom + vec2<f32>(0.5, 0.5) + pan;

    let v = sample_v(uv);

    // Central-difference gradient of the field (for hatch orientation + edges).
    let gx = sample_v(uv + vec2<f32>(texel.x, 0.0)) - sample_v(uv - vec2<f32>(texel.x, 0.0));
    let gy = sample_v(uv + vec2<f32>(0.0, texel.y)) - sample_v(uv - vec2<f32>(0.0, texel.y));
    let grad = vec2<f32>(gx, gy);
    let gmag = length(grad);

    let hue = pp.a.x;
    let density = pp.a.y;
    let hatch_freq = pp.a.z;
    let glow = pp.a.w;
    let color_span = pp.b.x;
    let color_center = pp.b.y;
    let saturation = pp.b.z;
    let palette_mix = pp.b.w;

    // Slope mask: contours and hatch only appear where the field actually
    // slopes, so the flat V=0 background stays dark (V=0 is itself an iso-level,
    // which would otherwise flood the flats).
    let slope = smoothstep(0.0008, 0.004, gmag);

    // Iso-contour lines: distance (in pixels) to the nearest V = k/density level,
    // anti-aliased by fwidth. `contour` is ~1 on a line, 0 between them.
    let f = v * density;
    let line_d = abs(fract(f - 0.5) - 0.5) / max(fwidth(f), 1e-4);
    let contour = (1.0 - clamp(line_d, 0.0, 1.0)) * slope;

    // Palette by field level so the nested loops read as coloured bands. The
    // field level `v` is the gradient coordinate: `color_span` (was a fixed 0.85)
    // sets the spanned range, `color_center`/`hue` slide the window, and the A/B
    // LUTs crossfade by `palette_mix` before the shared `saturation`.
    let coord = v * color_span + color_center + hue;
    // Hard bands, then the contour from the SAME coordinate (ADR-0078), so the
    // dark line follows the palette's iso-lines through the field.
    let banded = band_coord(coord, pp.d.x);
    let ca = textureSample(lut_a, lut_samp, vec2<f32>(banded, 0.5)).rgb;
    let cb = textureSample(lut_b, lut_samp, vec2<f32>(banded, 0.5)).rgb;
    let mixed = band_contour_ink(
        mix(ca, cb, clamp(palette_mix, 0.0, 1.0)),
        coord, pp.d.x, pp.d.y, pp.d.z, pp.d.w, lut_a, lut_b, lut_samp, palette_mix
    );
    let col = apply_saturation(mixed, saturation);

    // Hatch/comb: stripes along the contour tangent (perpendicular to grad),
    // gated to the slopes so flats stay clean.
    let tang = normalize(vec2<f32>(-grad.y, grad.x) + vec2<f32>(1e-5, 1e-5));
    let s = dot(uv * dims, tang) / max(hatch_freq, 1.0);
    let hatch = smoothstep(0.30, 0.5, abs(fract(s) - 0.5));
    let hatch_amt = hatch * slope;

    // Compose: dark bed, a coloured fill only where the field lives, bright
    // contour loops, hatch ticks that darken along the slopes, and a soft glow.
    let structure = smoothstep(0.04, 0.45, v);
    var out_col = col * structure * 0.5;
    out_col = out_col + col * contour * 0.9;
    out_col = out_col * (1.0 - hatch_amt * 0.4);
    out_col = out_col + col * v * glow * 0.22;

    // Alpha carries scene presence (the V-field `structure` term) so V=0 voids are
    // transparent and the `bg_*` backdrop shows through (ADR-0026). The present
    // pipeline blends premultiplied-OVER: `out_col` is emitted as-is (added over the
    // backdrop, so bright contours keep full brightness), and alpha only gates how
    // much backdrop reveals. Over the default black backdrop this is byte-identical
    // to the prior opaque present.
    //
    // `occlude` (pp.c.w) scales that coverage: how much of the backdrop the field
    // holds out where it does have presence (ADR-0085). Reached only when no post
    // stage is active — the chain's last stage owns the seam otherwise, and the
    // renderer hands a literal 1.0 here.
    return vec4<f32>(out_col, structure * pp.c.w);
}
"#;

#[repr(C)]
#[derive(Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
struct InitParams {
    blobs: [[f32; 4]; MAX_BLOBS],
    count: [u32; 4],
}

#[repr(C)]
#[derive(Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
struct SimParams {
    /// x: feed, y: kill, z: diffuse_u, w: diffuse_v.
    p: [f32; 4],
    /// xy: injection stamp center (uv), z: radius, w: amount (0 = none).
    inj: [f32; 4],
}

#[repr(C)]
#[derive(Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
struct PresentParams {
    /// x: hue, y: contour density, z: hatch frequency (texels), w: glow.
    a: [f32; 4],
    /// x: color_span, y: color_center, z: saturation, w: palette_mix.
    b: [f32; 4],
    /// x: zoom, yz: pan (view transform, ADR-0018), w: occlude (ADR-0085).
    c: [f32; 4],
    /// x: palette_steps, y: palette_contour (ADR-0078), z: palette_contour_style,
    /// w: palette_contour_ink (ADR-0197).
    d: [f32; 4],
}

/// The GPU-side state, built lazily on first render (see the module docs).
struct Resources {
    field: PingPongField,
    sim_pipeline: wgpu::RenderPipeline,
    init_pipeline: wgpu::RenderPipeline,
    present_pipeline: wgpu::RenderPipeline,
    sim_uniform: wgpu::Buffer,
    init_uniform: wgpu::Buffer,
    present_uniform: wgpu::Buffer,
    /// Sim/present bind groups reading texture A / texture B — selected by the
    /// field's read side each sub-step so nothing is rebuilt on the hot path.
    sim_bg_a: wgpu::BindGroup,
    sim_bg_b: wgpu::BindGroup,
    init_bg: wgpu::BindGroup,
    present_bg_a: wgpu::BindGroup,
    present_bg_b: wgpu::BindGroup,
    /// The shared gradient LUT pair (A/B) the present pass samples + crossfades
    /// (ADR-0021). A fresh pair is dirty, so a (re)build uploads on its first
    /// frame; a preset switch re-`set`s it through the scene's `set_palette`.
    luts: palette::LutPair,
}

impl Resources {
    /// Create every pipeline, buffer, bind group, and the ping-pong field.
    fn build(device: &wgpu::Device, surface_format: wgpu::TextureFormat) -> Self {
        let init_shader = gpu::fullscreen_shader(
            device,
            "rd-init-shader",
            gpu::FULLSCREEN_VS_UV_FLIPPED,
            INIT_SHADER,
        );
        let sim_shader = gpu::fullscreen_shader(
            device,
            "rd-sim-shader",
            gpu::FULLSCREEN_VS_UV_FLIPPED,
            SIM_SHADER,
        );
        let present_shader = gpu::fullscreen_shader(
            device,
            "rd-present-shader",
            gpu::FULLSCREEN_VS_UV_FLIPPED,
            PRESENT_SHADER,
        );

        let field = PingPongField::new(device, GRID, GRID);

        let sim_uniform =
            gpu::uniform_buffer(device, "rd-sim-params", std::mem::size_of::<SimParams>());
        let init_uniform =
            gpu::uniform_buffer(device, "rd-init-params", std::mem::size_of::<InitParams>());
        let present_uniform = gpu::uniform_buffer(
            device,
            "rd-present-params",
            std::mem::size_of::<PresentParams>(),
        );
        // --- init pipeline: one uniform, writes the seed field ---
        let init_layout = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
            label: Some("rd-init-layout"),
            entries: &[gpu::uniform(0, wgpu::ShaderStages::FRAGMENT)],
        });
        let init_bg = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rd-init-bg"),
            layout: &init_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: init_uniform.as_entire_binding(),
            }],
        });
        let init_pipeline = gpu::fullscreen_pipeline(
            device,
            &init_shader,
            &[&init_layout],
            PingPongField::FORMAT,
            wgpu::BlendState::REPLACE,
            "rd-init",
        );

        // --- sim pipeline: uniform + input texture (textureLoad, no sampler) ---
        let sim_layout = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
            label: Some("rd-sim-layout"),
            entries: &[
                gpu::uniform(0, wgpu::ShaderStages::FRAGMENT),
                gpu::texture(1, false),
            ],
        });
        let sim_bg_a = sim_bind_group(device, &sim_layout, &sim_uniform, field.view_a());
        let sim_bg_b = sim_bind_group(device, &sim_layout, &sim_uniform, field.view_b());
        let sim_pipeline = gpu::fullscreen_pipeline(
            device,
            &sim_shader,
            &[&sim_layout],
            PingPongField::FORMAT,
            wgpu::BlendState::REPLACE,
            "rd-sim",
        );

        // --- present pipeline: input texture + filtering sampler, to surface ---
        //
        // `Repeat`, not `ClampToEdge` (Plan 0033 Phase 5, ADR-0034). The
        // simulation has always been **toroidal** — `ld()` in the sim shader wraps
        // with `((c % size) + size) % size`, so the field is seamless — and only
        // the present sampler refused to wrap. Clamping smeared the edge row
        // outward into vertical bars and rectangular blocks the moment `zoom`
        // went above 1, and any real `pan_*` walked off the field, which is why
        // all four RD presets were pinned at `zoom = 0.99` and lost the whole
        // view lever. One address mode restores it: `pan_*` becomes a seamless
        // infinite scroll and `zoom > 1` tiles.
        //
        // It also makes the Catmull-Rom taps correct at the boundary: the p0/p3
        // taps reach one texel outside the cell, and wrapping is what the
        // toroidal field actually means there.
        let sampler = device.create_sampler(&wgpu::SamplerDescriptor {
            label: Some("rd-present-sampler"),
            address_mode_u: wgpu::AddressMode::Repeat,
            address_mode_v: wgpu::AddressMode::Repeat,
            address_mode_w: wgpu::AddressMode::Repeat,
            mag_filter: wgpu::FilterMode::Linear,
            min_filter: wgpu::FilterMode::Linear,
            ..Default::default()
        });
        // Shared gradient LUTs (ADR-0021): two 256×1 textures (A/B) + a repeat
        // sampler, in the present bind group alongside the field.
        let luts = palette::LutPair::new(device, "rd");
        let present_layout = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
            label: Some("rd-present-layout"),
            entries: &[
                gpu::texture(0, true),
                gpu::sampler(1),
                gpu::uniform(2, wgpu::ShaderStages::FRAGMENT),
                gpu::texture(3, true),
                gpu::texture(4, true),
                gpu::sampler(5),
            ],
        });
        let present_bg_a = present_bind_group(
            device,
            &present_layout,
            field.view_a(),
            &sampler,
            &present_uniform,
            &luts,
        );
        let present_bg_b = present_bind_group(
            device,
            &present_layout,
            field.view_b(),
            &sampler,
            &present_uniform,
            &luts,
        );
        let present_pipeline = gpu::fullscreen_pipeline(
            device,
            &present_shader,
            &[&present_layout],
            surface_format,
            // Premultiplied-alpha OVER the backdrop (ADR-0026): the scene is
            // emissive, so `out_col` adds over the atmosphere and the present's
            // alpha (scene presence) reveals `bg_*` in the field's voids. Over
            // the default black backdrop this equals the prior opaque REPLACE.
            wgpu::BlendState::PREMULTIPLIED_ALPHA_BLENDING,
            "rd-present",
        );

        Self {
            field,
            sim_pipeline,
            init_pipeline,
            present_pipeline,
            sim_uniform,
            init_uniform,
            present_uniform,
            sim_bg_a,
            sim_bg_b,
            init_bg,
            present_bg_a,
            present_bg_b,
            luts,
        }
    }

    /// Encode the one-shot seed pass into the current read texture, filling the
    /// field with the deterministic initial pattern. Run once after a (re)build.
    fn encode_seed(&self, encoder: &mut wgpu::CommandEncoder) {
        let mut pass = gpu::color_pass(
            encoder,
            "rd-seed-pass",
            self.field.read_view(),
            wgpu::LoadOp::Clear(wgpu::Color::BLACK),
        );
        pass.set_pipeline(&self.init_pipeline);
        pass.set_bind_group(0, &self.init_bg, &[]);
        pass.draw(0..3, 0..1);
    }
}

/// Gray-Scott reaction-diffusion on a ping-pong field, driven by named preset
/// parameters (ADR-0002 layer 2): `feed`/`kill` pick the regime, `flow` scales
/// the diffusion, and a rising `inject` edge stamps a seeded blob of growth.
pub struct ReactionDiffusionScene {
    /// Cloned device handle (an `Arc` inside wgpu) that builds
    /// [`Resources`] lazily on first render — see the module docs for
    /// why.
    device: wgpu::Device,
    surface_format: wgpu::TextureFormat,
    res: Option<Resources>,
    /// The deterministic seed pattern, uploaded on the first frame after a
    /// (re)build so a rebuilt scene restarts identically (capture determinism).
    init_params: InitParams,
    needs_seed: bool,
    /// Fixed-timestep accumulator: unspent injected `dt`, drained one
    /// [`FIXED_STEP`] at a time in [`advance`](Scene::advance).
    fixed_step: gpu::FixedStep,
    /// Sub-steps `advance` scheduled for the next `render` to encode.
    pending_substeps: u32,
    /// Seeded RNG for injection stamp positions (NFR 6); advanced only when a
    /// stamp fires, and reset with the scene so a capture reproduces exactly.
    stamp_rng: SeededRng,
    /// A stamp scheduled by an `inject` rising edge for the next `render`:
    /// (cx, cy, radius, amount). `None` when no beat fired this frame.
    pending_stamp: Option<[f32; 4]>,
    /// Previous frame's `inject` value, for rising-edge detection.
    prev_inject: f32,
    /// Shared scene clock (seconds), set by the renderer each frame.
    time: f32,
    feed: f32,
    kill: f32,
    /// Diffusion scale (multiplies both species' rates, keeping their ratio).
    flow: f32,
    /// This frame's injection level (bound to a beat/onset expression).
    inject: f32,
    /// The shared palette knobs (ADR-0021). This scene has no `brightness`.
    colour: common::PaletteParams,
    /// The shared view transform (ADR-0018).
    pan: common::PanParams,
    contour: f32,
    hatch: f32,
    glow: f32,
    /// Shared palette color knobs (ADR-0021 / Plan 0020 Phase 5).
    color_span: f32,
    color_center: f32,
    /// Shared view transform (ADR-0018 / Plan 0025 Phase 2): `zoom` scales the
    /// present-pass sample window about its centre, `pan_*` offsets it.
    zoom: f32,
    /// How much of this field's coverage the backdrop resolves against
    /// (ADR-0085). Set by the renderer every frame through
    /// [`Scene::set_occlude`](super::Scene::set_occlude) — not a named param, so
    /// `reset_params` leaves it alone.
    occlude: f32,
    /// The active baked palette. Held here rather than only in the resources'
    /// [`palette::LutPair`] because the resources build lazily: `set_palette` can
    /// arrive with `res` still `None`, and this is what seeds the pair when it
    /// finally exists.
    palette: Palette,
}

impl ReactionDiffusionScene {
    /// Build the CPU-side state and compute the deterministic seed pattern. GPU
    /// resources are deferred to the first render (module docs).
    pub fn new(device: &wgpu::Device, surface_format: wgpu::TextureFormat) -> Self {
        let mut init_params = InitParams {
            blobs: [[0.0; 4]; MAX_BLOBS],
            count: [0; 4],
        };
        let mut rng = SeededRng::new(SEED);
        let n = SEED_BLOBS.min(MAX_BLOBS);
        for slot in init_params.blobs.iter_mut().take(n) {
            let x = rng.next_f32();
            let y = rng.next_f32();
            let r = rng.range(0.02, 0.045);
            *slot = [x, y, r, 0.0];
        }
        init_params.count = [n as u32, 0, 0, 0];

        Self {
            device: device.clone(),
            surface_format,
            res: None,
            init_params,
            needs_seed: true,
            fixed_step: gpu::FixedStep::new(FIXED_STEP, MAX_SUBSTEPS),
            pending_substeps: 0,
            stamp_rng: SeededRng::new(INJECT_SEED),
            pending_stamp: None,
            prev_inject: 0.0,
            time: 0.0,
            feed: DEFAULT_FEED,
            kill: DEFAULT_KILL,
            flow: DEFAULT_FLOW,
            inject: 0.0,
            colour: common::PaletteParams::new(DEFAULT_HUE, common::DEFAULT_BRIGHTNESS),
            pan: common::PanParams::default(),
            contour: DEFAULT_CONTOUR,
            hatch: DEFAULT_HATCH,
            glow: DEFAULT_GLOW,
            color_span: DEFAULT_COLOR_SPAN,
            color_center: DEFAULT_COLOR_CENTER,
            zoom: DEFAULT_ZOOM,
            occlude: crate::render::post::DEFAULT_OCCLUDE,
            palette: Palette::default_spectrum(),
        }
    }
}

fn sim_bind_group(
    device: &wgpu::Device,
    layout: &wgpu::BindGroupLayout,
    uniform: &wgpu::Buffer,
    input: &wgpu::TextureView,
) -> wgpu::BindGroup {
    device.create_bind_group(&wgpu::BindGroupDescriptor {
        label: Some("rd-sim-bg"),
        layout,
        entries: &[
            wgpu::BindGroupEntry {
                binding: 0,
                resource: uniform.as_entire_binding(),
            },
            wgpu::BindGroupEntry {
                binding: 1,
                resource: wgpu::BindingResource::TextureView(input),
            },
        ],
    })
}

fn present_bind_group(
    device: &wgpu::Device,
    layout: &wgpu::BindGroupLayout,
    input: &wgpu::TextureView,
    sampler: &wgpu::Sampler,
    uniform: &wgpu::Buffer,
    luts: &palette::LutPair,
) -> wgpu::BindGroup {
    let [lut_a, lut_b, lut_sampler] = luts.bind_entries(3, 4, 5);
    device.create_bind_group(&wgpu::BindGroupDescriptor {
        label: Some("rd-present-bg"),
        layout,
        entries: &[
            wgpu::BindGroupEntry {
                binding: 0,
                resource: wgpu::BindingResource::TextureView(input),
            },
            wgpu::BindGroupEntry {
                binding: 1,
                resource: wgpu::BindingResource::Sampler(sampler),
            },
            wgpu::BindGroupEntry {
                binding: 2,
                resource: uniform.as_entire_binding(),
            },
            lut_a,
            lut_b,
            lut_sampler,
        ],
    })
}

/// Parameter vocabulary — see [`fragment_field::PARAMS`](super::fragment_field::PARAMS).
/// **Keep in sync with `set_param` below.**
pub const PARAMS: &[ParamSpec] = &[
    ParamSpec {
        name: "feed",
        default: 0.0367,
        range: Some([0.01, 0.09]),
        doc: "Feed rate of the reaction - with `kill`, it is what decides whether you get spots, stripes or mitosis.",
        kind: ParamKind::Modal,
        group: ParamGroup::Shape,
        main: true,
    },
    ParamSpec {
        name: "kill",
        default: 0.0649,
        range: Some([0.03, 0.07]),
        doc: "Kill rate of the reaction; small moves here change the pattern's whole character.",
        kind: ParamKind::Modal,
        group: ParamGroup::Shape,
        main: true,
    },
    ParamSpec {
        name: "flow",
        default: 1.0,
        range: Some([0.0, 4.0]),
        doc: "How fast the simulation advances per second.",
        kind: ParamKind::Modal,
        group: ParamGroup::Motion,
        main: true,
    },
    ParamSpec {
        name: "inject",
        default: 0.0,
        range: Some([0.0, 1.0]),
        doc: "Drops fresh reagent into the field, which is how a beat seeds new growth.",
        kind: ParamKind::Modal,
        group: ParamGroup::Motion,
        main: false,
    },
    crate::render::scenes::common::hue(DEFAULT_HUE),
    ParamSpec {
        name: "contour",
        default: 6.0,
        range: Some([0.0, 24.0]),
        doc: "How many bands the concentration is drawn as, as a real density: a fraction \
               slides the whole set of iso-lines. 0 is a smooth gradient.",
        kind: ParamKind::Modal,
        group: ParamGroup::Shape,
        main: false,
    },
    ParamSpec {
        name: "hatch",
        default: 5.0,
        range: Some([0.0, 24.0]),
        doc: "Density of the hatching drawn along the concentration gradient.",
        kind: ParamKind::Modal,
        group: ParamGroup::Shape,
        main: false,
    },
    ParamSpec {
        name: "glow",
        default: 1.0,
        range: Some([0.0, 2.0]),
        doc: "Overall light the field emits.",
        kind: ParamKind::Modal,
        group: ParamGroup::Light,
        main: false,
    },
    ParamSpec {
        name: "color_span",
        default: 0.85,
        range: Some([0.0, 1.0]),
        doc: "How much of the palette the concentration range covers.",
        kind: ParamKind::Modal,
        group: ParamGroup::Colour,
        main: false,
    },
    ParamSpec {
        name: "color_center",
        default: 0.0,
        range: Some([-1.0, 1.0]),
        doc: "Shifts which concentration lands in the middle of the palette.",
        kind: ParamKind::Modal,
        group: ParamGroup::Colour,
        main: false,
    },
    crate::render::scenes::common::SATURATION,
    crate::render::scenes::common::PALETTE_MIX,
    crate::render::scenes::common::PALETTE_STEPS,
    crate::render::scenes::common::PALETTE_CONTOUR,
    crate::render::scenes::common::PALETTE_CONTOUR_STYLE,
    crate::render::scenes::common::PALETTE_CONTOUR_INK,
    crate::render::scenes::common::field_zoom(DEFAULT_ZOOM),
    crate::render::scenes::common::PAN_X,
    crate::render::scenes::common::PAN_Y,
];

impl Scene for ReactionDiffusionScene {
    fn name(&self) -> &'static str {
        "reaction diffusion"
    }

    fn advance(&mut self, dt: f32) {
        // Drain the accumulator one fixed sub-step at a time, clamped so a long
        // stall can't queue unbounded work (ADR-0012). The sub-`FIXED_STEP`
        // remainder carries to the next frame; a clamp drops the excess backlog
        // so the sim slows rather than races to catch up. Shared with the
        // attractor scene, which drains an identical accumulator.
        self.pending_substeps = self.fixed_step.advance(dt);
    }

    fn set_time(&mut self, time: f32) {
        self.time = time;
    }

    fn set_occlude(&mut self, occlude: f32) {
        self.occlude = occlude;
    }

    fn set_palette(&mut self, palette: &Palette) {
        // Uploaded to the present LUT textures in `render` (deferred — resources
        // build lazily on first render). Cheap array copy, off the hot path.
        self.palette = palette.clone();
        if let Some(res) = self.res.as_mut() {
            res.luts.set(palette);
        }
    }

    fn reset_params(&mut self) {
        self.feed = DEFAULT_FEED;
        self.kill = DEFAULT_KILL;
        self.flow = DEFAULT_FLOW;
        self.inject = 0.0;
        self.colour.reset();
        self.pan.reset();
        self.contour = DEFAULT_CONTOUR;
        self.hatch = DEFAULT_HATCH;
        self.glow = DEFAULT_GLOW;
        self.color_span = DEFAULT_COLOR_SPAN;
        self.color_center = DEFAULT_COLOR_CENTER;
        self.zoom = DEFAULT_ZOOM;
    }

    fn set_param(&mut self, name: &str, value: f32) {
        // The shared param blocks first, this scene's own names after
        // (`scenes::common`).
        if self.colour.set(name, value) || self.pan.set(name, value) {
            return;
        }
        // ADR-0002 layer 2 knobs. `feed`/`kill` pick the regime; `flow` scales
        // the diffusion; `inject` is a beat/onset level whose rising edge stamps
        // a seed (edge detected in `update`). `hue`/`contour`/`hatch`/`glow`
        // drive the iso-contour present look (Phase 4).
        match name {
            "feed" => self.feed = value,
            "kill" => self.kill = value,
            "flow" => self.flow = value,
            "inject" => self.inject = value,
            "contour" => self.contour = value,
            "hatch" => self.hatch = value,
            "glow" => self.glow = value,
            "color_span" => self.color_span = value,
            "color_center" => self.color_center = value,
            "zoom" => self.zoom = value,
            _ => {}
        }
    }

    fn update(&mut self, _frame: &AnalysisFrame) {
        // Rising-edge detect on `inject` (a beat/onset expression): schedule one
        // stamp at the next seeded position. Edge-triggered so a sustained beat
        // flag doesn't stamp every frame; deterministic because the position
        // comes from the seeded, capture-reset `stamp_rng` (NFR 6).
        if self.inject >= INJECT_THRESHOLD && self.prev_inject < INJECT_THRESHOLD {
            let cx = self.stamp_rng.next_f32();
            let cy = self.stamp_rng.next_f32();
            self.pending_stamp = Some([cx, cy, INJECT_RADIUS, INJECT_AMOUNT]);
        }
        self.prev_inject = self.inject;
    }

    fn render(
        &mut self,
        queue: &wgpu::Queue,
        encoder: &mut wgpu::CommandEncoder,
        view: &wgpu::TextureView,
        _aspect: f32,
    ) {
        // Build GPU resources on first use (module docs). A fresh pair's textures
        // are empty and hold the default palette, so hand it the one the scene is
        // actually carrying; the upload happens in the flush below.
        if self.res.is_none() {
            let mut built = Resources::build(&self.device, self.surface_format);
            built.luts.set(&self.palette);
            self.res = Some(built);
        }
        let Self {
            res,
            init_params,
            needs_seed,
            pending_substeps,
            pending_stamp,
            feed,
            kill,
            flow,
            contour,
            hatch,
            glow,
            color_span,
            color_center,
            colour,
            zoom,
            pan,
            occlude,
            ..
        } = self;
        let Some(res) = res.as_mut() else {
            return;
        };

        // Upload the active palette LUTs (A + B) on a preset switch or a fresh
        // build — off the hot path, once per change.
        res.luts.flush(queue);

        queue.write_buffer(
            &res.present_uniform,
            0,
            bytemuck::bytes_of(&PresentParams {
                a: [colour.hue, *contour, *hatch, *glow],
                b: [*color_span, *color_center, colour.saturation, colour.mix],
                c: [*zoom, pan.x, pan.y, *occlude],
                d: [
                    palette::band_steps(colour.steps),
                    palette::band_contour(colour.contour),
                    palette::band_contour_style(colour.contour_style),
                    colour.contour_ink,
                ],
            }),
        );

        // A beat scheduled a stamp this frame (consumed here): the sim shader
        // applies it on every sub-step, so V saturates at the stamp. `[0; 4]`
        // means no injection.
        let inj = pending_stamp.take().unwrap_or([0.0; 4]);
        queue.write_buffer(
            &res.sim_uniform,
            0,
            bytemuck::bytes_of(&SimParams {
                // `flow` scales both diffusion rates, keeping the 2:1 ratio (so
                // the pattern coarsens/tightens without changing regime).
                p: [*feed, *kill, DIFFUSE_U * *flow, DIFFUSE_V * *flow],
                inj,
            }),
        );

        // One-shot deterministic seed on the first frame after a (re)build.
        if *needs_seed {
            queue.write_buffer(&res.init_uniform, 0, bytemuck::bytes_of(init_params));
            res.encode_seed(encoder);
            *needs_seed = false;
        }

        // Run the sub-steps the accumulator scheduled this frame (Phase 2). Each
        // reads the current field and writes the other texture, then swaps.
        for _ in 0..*pending_substeps {
            let sim_bg = if res.field.reading_a() {
                &res.sim_bg_a
            } else {
                &res.sim_bg_b
            };
            {
                // The fullscreen sim pass overwrites every texel.
                let mut pass = gpu::color_pass(
                    encoder,
                    "rd-sim-pass",
                    res.field.write_view(),
                    wgpu::LoadOp::Clear(wgpu::Color::BLACK),
                );
                pass.set_pipeline(&res.sim_pipeline);
                pass.set_bind_group(0, sim_bg, &[]);
                pass.draw(0..3, 0..1);
            }
            res.field.swap();
        }

        // Present the latest field to the surface.
        let present_bg = if res.field.reading_a() {
            &res.present_bg_a
        } else {
            &res.present_bg_b
        };
        // Load over the engine backdrop (ADR-0018). The present is **not**
        // opaque: it writes premultiplied alpha, with the V-field's `structure`
        // term as coverage, so the coral's voids reveal the `bg_*` gradient
        // underneath. Over the default black backdrop a premultiplied Load is
        // arithmetically an opaque REPLACE, which is why the golden does not
        // move.
        let mut pass = gpu::color_pass(encoder, "rd-present-pass", view, wgpu::LoadOp::Load);
        pass.set_pipeline(&res.present_pipeline);
        pass.set_bind_group(0, present_bg, &[]);
        pass.draw(0..3, 0..1);
    }
}