From 42635eeeff5e84e186730348d1d4660d236bfab2 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Fri, 25 Sep 2026 17:42:29 +0000
Subject: [PATCH 1/5] MONKEY (polish): sun shafts march a half-resolution sky
 mask

The radial blur read the scene depth 28 times per full-resolution pixel. A
new mask pre-pass (`fs_mask`, R8Unorm at ceil(w/2) x ceil(h/2)) loads the
depth once per half-res texel and thresholds it with the same sky test
(depth <= 1e-7 on infinite reverse-Z); the blur then takes 28 clamp-to-edge
bilinear taps of the mask. Only the mask stage is MSAA-specialised (single
and multisampled depth layouts, one pipeline each); the blur pipeline is
keyed by the target format alone.

Verified: cargo check -p benilla; all four WGSL variants (blur, blur +
MULTISAMPLED, SHAFT_MASK, SHAFT_MASK + MULTISAMPLED) compose and validate
under naga_oil 0.20 / naga 27 offline. No GPU or WoW data in this
environment, so the image is not captured; built on Linux only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y6yUWgPosSakvL23gtgvpH
---
 crates/benilla-app/src/post/sun_shafts.rs   | 182 ++++++++++++++++----
 crates/benilla-app/src/post/sun_shafts.wgsl |  49 +++---
 2 files changed, 172 insertions(+), 59 deletions(-)

diff --git a/crates/benilla-app/src/post/sun_shafts.rs b/crates/benilla-app/src/post/sun_shafts.rs
index f509fd48..04228740 100644
--- a/crates/benilla-app/src/post/sun_shafts.rs
+++ b/crates/benilla-app/src/post/sun_shafts.rs
@@ -1,4 +1,7 @@
 //! MONKEY (post): screen-space sun shafts from the visible sky depth.
+//!
+//! MONKEY (polish): the depth is read once per half-resolution texel into an R8 sky mask
+//! (`fs_mask`, the only MSAA-specialised stage); the 28-tap radial blur then samples the mask.
 
 use super::bloom::BloomLabel;
 use crate::video::VideoConfig;
@@ -11,6 +14,7 @@ use bevy::{
     ecs::query::QueryItem,
     prelude::*,
     render::{
+        camera::ExtractedCamera,
         diagnostic::RecordDiagnostics,
         extract_component::{ExtractComponent, ExtractComponentPlugin},
         render_graph::{
@@ -18,7 +22,8 @@ use bevy::{
         },
         render_resource::{binding_types::*, *},
         renderer::{RenderContext, RenderDevice, RenderQueue},
-        view::{Msaa, ViewDepthTexture, ViewTarget},
+        texture::{CachedTexture, TextureCache},
+        view::{ViewDepthTexture, ViewTarget},
         Render, RenderApp, RenderStartup, RenderSystems,
     },
 };
@@ -54,7 +59,13 @@ impl Plugin for SunShaftsPlugin {
             .insert_resource(ShaftShader(shader))
             .init_resource::<SpecializedRenderPipelines<ShaftPipeline>>()
             .add_systems(RenderStartup, init_pipeline)
-            .add_systems(Render, prepare_pipelines.in_set(RenderSystems::Prepare))
+            .add_systems(
+                Render,
+                (
+                    prepare_pipelines.in_set(RenderSystems::Prepare),
+                    prepare_masks.in_set(RenderSystems::PrepareResources),
+                ),
+            )
             .add_render_graph_node::<ViewNodeRunner<ShaftNode>>(Core3d, ShaftLabel)
             .add_render_graph_edges(
                 Core3d,
@@ -139,12 +150,21 @@ struct ShaftShader(Handle<Shader>);
 
 #[derive(Resource)]
 struct ShaftPipeline {
-    layouts: [BindGroupLayoutDescriptor; 2],
+    /// The blur's layout; the mask's per MSAA (`[single, multisampled]`).
+    layout: BindGroupLayoutDescriptor,
+    mask_layouts: [BindGroupLayoutDescriptor; 2],
+    masks: [CachedRenderPipelineId; 2],
     shader: Handle<Shader>,
     fullscreen: FullscreenShader,
     sampler: Sampler,
 }
 
+/// The half-resolution sky mask the blur marches through.
+const MASK_FORMAT: TextureFormat = TextureFormat::R8Unorm;
+
+#[derive(Component)]
+struct ShaftMask(CachedTexture);
+
 #[derive(Component)]
 struct ViewShaftPipeline(CachedRenderPipelineId);
 
@@ -154,29 +174,61 @@ struct ShaftLabel;
 fn init_pipeline(
     mut commands: Commands,
     device: Res<RenderDevice>,
+    cache: Res<PipelineCache>,
     shader: Res<ShaftShader>,
     fullscreen: Res<FullscreenShader>,
 ) {
-    let layouts = [false, true].map(|multisampled| {
+    let layout = BindGroupLayoutDescriptor::new(
+        "post_sun_shafts_layout",
+        &BindGroupLayoutEntries::sequential(
+            ShaderStages::FRAGMENT,
+            (
+                texture_2d(TextureSampleType::Float { filterable: true }),
+                sampler(SamplerBindingType::Filtering),
+                texture_2d(TextureSampleType::Float { filterable: true }),
+                uniform_buffer::<ShaftView>(false),
+            ),
+        ),
+    );
+    let mask_layouts = [false, true].map(|multisampled| {
         BindGroupLayoutDescriptor::new(
-            "post_sun_shafts_layout",
-            &BindGroupLayoutEntries::sequential(
+            "post_sun_shafts_mask_layout",
+            &BindGroupLayoutEntries::single(
                 ShaderStages::FRAGMENT,
-                (
-                    texture_2d(TextureSampleType::Float { filterable: true }),
-                    sampler(SamplerBindingType::Filtering),
-                    if multisampled {
-                        texture_depth_2d_multisampled()
-                    } else {
-                        texture_depth_2d()
-                    },
-                    uniform_buffer::<ShaftView>(false),
-                ),
+                if multisampled {
+                    texture_depth_2d_multisampled()
+                } else {
+                    texture_depth_2d()
+                },
             ),
         )
     });
+    let masks = [false, true].map(|multisampled| {
+        let mut shader_defs = vec!["SHAFT_MASK".into()];
+        if multisampled {
+            shader_defs.push("MULTISAMPLED".into());
+        }
+        cache.queue_render_pipeline(RenderPipelineDescriptor {
+            label: Some("post_sun_shafts_mask".into()),
+            layout: vec![mask_layouts[multisampled as usize].clone()],
+            vertex: fullscreen.to_vertex_state(),
+            fragment: Some(FragmentState {
+                shader: shader.0.clone(),
+                shader_defs,
+                entry_point: Some("fs_mask".into()),
+                targets: vec![Some(ColorTargetState {
+                    format: MASK_FORMAT,
+                    blend: None,
+                    write_mask: ColorWrites::ALL,
+                })],
+            }),
+            ..default()
+        })
+    });
     commands.insert_resource(ShaftPipeline {
-        layouts,
+        layout,
+        mask_layouts,
+        masks,
         shader: shader.0.clone(),
         fullscreen: fullscreen.clone(),
         sampler: device.create_sampler(&SamplerDescriptor {
@@ -190,19 +242,16 @@ fn init_pipeline(
 }
 
 impl SpecializedRenderPipeline for ShaftPipeline {
-    type Key = (TextureFormat, bool);
+    type Key = TextureFormat;
 
-    fn specialize(&self, (format, multisampled): Self::Key) -> RenderPipelineDescriptor {
+    fn specialize(&self, format: Self::Key) -> RenderPipelineDescriptor {
         RenderPipelineDescriptor {
             label: Some("post_sun_shafts".into()),
-            layout: vec![self.layouts[multisampled as usize].clone()],
+            layout: vec![self.layout.clone()],
             vertex: self.fullscreen.to_vertex_state(),
             fragment: Some(FragmentState {
                 shader: self.shader.clone(),
-                shader_defs: multisampled
-                    .then(|| "MULTISAMPLED".into())
-                    .into_iter()
-                    .collect(),
+                shader_defs: vec![],
                 entry_point: Some("fragment".into()),
                 targets: vec![Some(ColorTargetState {
                     format,
@@ -220,18 +269,46 @@ fn prepare_pipelines(
     cache: Res<PipelineCache>,
     pipeline: Res<ShaftPipeline>,
     mut specialized: ResMut<SpecializedRenderPipelines<ShaftPipeline>>,
-    views: Query<(Entity, &ViewTarget, &Msaa), With<ShaftView>>,
+    views: Query<(Entity, &ViewTarget), With<ShaftView>>,
 ) {
-    for (entity, target, msaa) in &views {
-        let id = specialized.specialize(
-            &cache,
-            &pipeline,
-            (target.main_texture_format(), msaa.samples() > 1),
-        );
+    for (entity, target) in &views {
+        let id = specialized.specialize(&cache, &pipeline, target.main_texture_format());
         commands.entity(entity).insert(ViewShaftPipeline(id));
     }
 }
 
+/// The half-size mask target, rounded up so an odd edge column still has a texel.
+fn prepare_masks(
+    mut commands: Commands,
+    mut textures: ResMut<TextureCache>,
+    device: Res<RenderDevice>,
+    views: Query<(Entity, &ExtractedCamera), With<ShaftView>>,
+) {
+    for (entity, camera) in &views {
+        let Some(size) = camera.physical_viewport_size else {
+            continue;
+        };
+        let mask = textures.get(
+            &device,
+            TextureDescriptor {
+                label: Some("post_sun_shafts_mask"),
+                size: Extent3d {
+                    width: size.x.div_ceil(2).max(1),
+                    height: size.y.div_ceil(2).max(1),
+                    depth_or_array_layers: 1,
+                },
+                mip_level_count: 1,
+                sample_count: 1,
+                dimension: TextureDimension::D2,
+                format: MASK_FORMAT,
+                usage: TextureUsages::RENDER_ATTACHMENT | TextureUsages::TEXTURE_BINDING,
+                view_formats: &[],
+            },
+        );
+        commands.entity(entity).insert(ShaftMask(mask));
+    }
+}
+
 #[derive(Default)]
 struct ShaftNode;
 
@@ -241,17 +318,23 @@ impl ViewNode for ShaftNode {
         &'static ViewDepthTexture,
         &'static ShaftView,
         &'static ViewShaftPipeline,
+        &'static ShaftMask,
     );
 
     fn run<'w>(
         &self,
         _graph: &mut RenderGraphContext,
         context: &mut RenderContext<'w>,
-        (target, depth, shaft, id): QueryItem<'w, '_, Self::ViewQuery>,
+        (target, depth, shaft, id, mask): QueryItem<'w, '_, Self::ViewQuery>,
         world: &'w World,
     ) -> Result<(), NodeRunError> {
         let cache = world.resource::<PipelineCache>();
-        let Some(pipeline) = cache.get_render_pipeline(id.0) else {
+        let settings = world.resource::<ShaftPipeline>();
+        let multisampled = (depth.texture.sample_count() > 1) as usize;
+        let (Some(pipeline), Some(mask_pipeline)) = (
+            cache.get_render_pipeline(id.0),
+            cache.get_render_pipeline(settings.masks[multisampled]),
+        ) else {
             return Ok(());
         };
         if !depth
@@ -261,24 +344,47 @@ impl ViewNode for ShaftNode {
         {
             return Ok(());
         }
-        let settings = world.resource::<ShaftPipeline>();
         let device = context.render_device();
+        let mask_bind = device.create_bind_group(
+            "post_sun_shafts_mask",
+            &cache.get_bind_group_layout(&settings.mask_layouts[multisampled]),
+            &BindGroupEntries::single(depth.view()),
+        );
         let mut uniform = UniformBuffer::from(*shaft);
         uniform.write_buffer(device, world.resource::<RenderQueue>());
         let out = target.post_process_write();
         let bind = device.create_bind_group(
             "post_sun_shafts",
-            &cache.get_bind_group_layout(
-                &settings.layouts[(depth.texture.sample_count() > 1) as usize],
-            ),
+            &cache.get_bind_group_layout(&settings.layout),
             &BindGroupEntries::sequential((
                 out.source,
                 &settings.sampler,
-                depth.view(),
+                &mask.0.default_view,
                 uniform.binding().unwrap(),
             )),
         );
         let diagnostics = context.diagnostic_recorder();
+        {
+            let mut pass = context
+                .command_encoder()
+                .begin_render_pass(&RenderPassDescriptor {
+                    label: Some("post_sun_shafts_mask"),
+                    color_attachments: &[Some(RenderPassColorAttachment {
+                        view: &mask.0.default_view,
+                        depth_slice: None,
+                        resolve_target: None,
+                        ops: Operations::default(),
+                    })],
+                    depth_stencil_attachment: None,
+                    timestamp_writes: None,
+                    occlusion_query_set: None,
+                });
+            let span = diagnostics.pass_span(&mut pass, "post_sun_shafts_mask");
+            pass.set_pipeline(mask_pipeline);
+            pass.set_bind_group(0, &mask_bind, &[]);
+            pass.draw(0..3, 0..1);
+            span.end(&mut pass);
+        }
         let mut pass = context
             .command_encoder()
             .begin_render_pass(&RenderPassDescriptor {
diff --git a/crates/benilla-app/src/post/sun_shafts.wgsl b/crates/benilla-app/src/post/sun_shafts.wgsl
index 2ddc1834..f7dec61f 100644
--- a/crates/benilla-app/src/post/sun_shafts.wgsl
+++ b/crates/benilla-app/src/post/sun_shafts.wgsl
@@ -1,14 +1,33 @@
 // MONKEY (post): screen-space radial blur through sky pixels in the scene depth.
+// MONKEY (polish): two passes. `SHAFT_MASK` reads the depth once per half-resolution texel into
+// a sky mask; the blur then samples that mask 28 times instead of the full-size depth.
 
 #import bevy_core_pipeline::fullscreen_vertex_shader::FullscreenVertexOutput
 
+#ifdef SHAFT_MASK
+
+#ifdef MULTISAMPLED
+@group(0) @binding(0) var depth: texture_depth_multisampled_2d;
+#else
+@group(0) @binding(0) var depth: texture_depth_2d;
+#endif
+
+@fragment
+fn fs_mask(in: FullscreenVertexOutput) -> @location(0) vec4<f32> {
+    let dims = vec2<i32>(textureDimensions(depth));
+    let p = min(vec2<i32>(in.position.xy) * 2, dims - vec2<i32>(1));
+    // MONKEY (fix-post): one sample, not a max over all of them: the march is a blur.
+    // Only the cleared / sky-pinned depth (exactly 0 on infinite reverse-Z) is sky; a ramp on
+    // `near / z` would move with the live `nearclip` cvar and pour shafts through far terrain.
+    let sky = select(0.0, 1.0, textureLoad(depth, p, 0) <= 1.0e-7);
+    return vec4(sky, 0.0, 0.0, 1.0);
+}
+
+#else
+
 @group(0) @binding(0) var scene: texture_2d<f32>;
 @group(0) @binding(1) var scene_sampler: sampler;
-#ifdef MULTISAMPLED
-@group(0) @binding(2) var depth: texture_depth_multisampled_2d;
-#else
-@group(0) @binding(2) var depth: texture_depth_2d;
-#endif
+@group(0) @binding(2) var mask: texture_2d<f32>;
 
 struct Shafts {
     sun: vec4<f32>,
@@ -16,18 +35,6 @@ struct Shafts {
 };
 @group(0) @binding(3) var<uniform> shafts: Shafts;
 
-fn scene_depth(uv: vec2<f32>) -> f32 {
-    let dims = textureDimensions(depth);
-    let p = vec2<i32>(clamp(uv, vec2(0.0), vec2(0.999999)) * vec2<f32>(dims));
-#ifdef MULTISAMPLED
-    // MONKEY (fix-post): one sample, not a max over all of them: the march is a blur, and the
-    // per-sample loop cost 28 x N loads per pixel (224 at 8x).
-    return textureLoad(depth, p, 0);
-#else
-    return textureLoad(depth, p, 0);
-#endif
-}
-
 @fragment
 fn fragment(in: FullscreenVertexOutput) -> @location(0) vec4<f32> {
     let base = textureSample(scene, scene_sampler, in.uv);
@@ -41,10 +48,8 @@ fn fragment(in: FullscreenVertexOutput) -> @location(0) vec4<f32> {
         if (i >= count) { break; }
         uv += step_uv;
         decay *= 0.965;
-        // MONKEY (fix-post): only the cleared / sky-pinned depth (exactly 0 on infinite reverse-Z)
-        // is sky. The old ramp read d = near / z, so it moved with the live `nearclip` cvar and
-        // let shafts pour through terrain beyond ~125 yd at nearclip 0.01.
-        let sky = select(0.0, 1.0, scene_depth(uv) <= 1.0e-7);
+        // The clamp-to-edge bilinear tap is the old clamped depth read, pre-thresholded.
+        let sky = textureSampleLevel(mask, scene_sampler, uv, 0.0).r;
         light += sky * decay;
         norm += decay;
     }
@@ -56,3 +61,5 @@ fn fragment(in: FullscreenVertexOutput) -> @location(0) vec4<f32> {
     let rays = light / max(norm, 0.0001) * radial * radial;
     return vec4(base.rgb + shafts.color.rgb * rays * shafts.sun.z * 0.16, base.a);
 }
+
+#endif

From 78d7e94433e81ee1befd3359f85793e64db5bc7d Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Fri, 25 Sep 2026 17:46:00 +0000
Subject: [PATCH 2/5] MONKEY (polish): per-view post uniforms written at
 prepare, not allocated per frame

Colour grading and sun shafts built a fresh `UniformBuffer` (a new GPU buffer)
inside their render-graph node every frame. Each view now owns one small
UNIFORM | COPY_DST buffer, created the first frame the effect is on and
rewritten in place with `RenderQueue::write_buffer` in
`RenderSystems::PrepareResources`, the same shape as bloom's prepare-time
uniform. The node only binds it. The pass still gates on the extracted view
component, so the kept buffer never runs a disabled pass.

Verified: cargo check -p benilla-app; rustfmt clean on both files. No GPU
here; built on Linux only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y6yUWgPosSakvL23gtgvpH
---
 crates/benilla-app/src/post/grading.rs    | 46 +++++++++++++++++++----
 crates/benilla-app/src/post/sun_shafts.rs | 32 ++++++++++++----
 2 files changed, 63 insertions(+), 15 deletions(-)

diff --git a/crates/benilla-app/src/post/grading.rs b/crates/benilla-app/src/post/grading.rs
index 5a4dc42a..b3fd694d 100644
--- a/crates/benilla-app/src/post/grading.rs
+++ b/crates/benilla-app/src/post/grading.rs
@@ -25,7 +25,7 @@ use bevy::{
         renderer::{RenderContext, RenderDevice, RenderQueue},
         texture::GpuImage,
         view::ViewTarget,
-        RenderApp, RenderStartup,
+        Render, RenderApp, RenderStartup, RenderSystems,
     },
 };
 use std::collections::{HashMap, HashSet};
@@ -73,6 +73,10 @@ impl Plugin for GradingPlugin {
         render_app
             .insert_resource(GradeShader(shader))
             .add_systems(RenderStartup, init_pipeline)
+            .add_systems(
+                Render,
+                prepare_uniforms.in_set(RenderSystems::PrepareResources),
+            )
             .add_render_graph_node::<ViewNodeRunner<GradeNode>>(Core3d, GradeLabel)
             .add_render_graph_edges(
                 Core3d,
@@ -344,17 +348,47 @@ fn init_pipeline(
     });
 }
 
+/// MONKEY (polish): the view's uniform, created once and rewritten in prepare, like bloom's.
+#[derive(Component)]
+struct GradeUniformBuffer(Buffer);
+
+fn prepare_uniforms(
+    mut commands: Commands,
+    device: Res<RenderDevice>,
+    queue: Res<RenderQueue>,
+    views: Query<(Entity, &GradeView, Option<&GradeUniformBuffer>)>,
+) {
+    for (entity, grade, buffer) in &views {
+        let rows = [grade.control.to_array()];
+        match buffer {
+            Some(buffer) => queue.write_buffer(&buffer.0, 0, bytemuck::cast_slice(&rows)),
+            None => {
+                let buffer = device.create_buffer_with_data(&BufferInitDescriptor {
+                    label: Some("post_grading_uniform"),
+                    contents: bytemuck::cast_slice(&rows),
+                    usage: BufferUsages::UNIFORM | BufferUsages::COPY_DST,
+                });
+                commands.entity(entity).insert(GradeUniformBuffer(buffer));
+            }
+        }
+    }
+}
+
 #[derive(Default)]
 struct GradeNode;
 
 impl ViewNode for GradeNode {
-    type ViewQuery = (&'static ViewTarget, &'static GradeView);
+    type ViewQuery = (
+        &'static ViewTarget,
+        &'static GradeView,
+        &'static GradeUniformBuffer,
+    );
 
     fn run<'w>(
         &self,
         _graph: &mut RenderGraphContext,
         context: &mut RenderContext<'w>,
-        (target, grade): QueryItem<'w, '_, Self::ViewQuery>,
+        (target, grade, uniform): QueryItem<'w, '_, Self::ViewQuery>,
         world: &'w World,
     ) -> Result<(), NodeRunError> {
         let settings = world.resource::<GradePipeline>();
@@ -367,10 +401,6 @@ impl ViewNode for GradeNode {
             return Ok(());
         };
         let device = context.render_device();
-        let mut uniform = UniformBuffer::from(GradeUniform {
-            control: grade.control,
-        });
-        uniform.write_buffer(device, world.resource::<RenderQueue>());
         let out = target.post_process_write();
         let bind = device.create_bind_group(
             "post_grading",
@@ -381,7 +411,7 @@ impl ViewNode for GradeNode {
                 &day.texture_view,
                 &night.texture_view,
                 &settings.lut_sampler,
-                uniform.binding().unwrap(),
+                uniform.0.as_entire_binding(),
             )),
         );
         let diagnostics = context.diagnostic_recorder();
diff --git a/crates/benilla-app/src/post/sun_shafts.rs b/crates/benilla-app/src/post/sun_shafts.rs
index 04228740..b82420d1 100644
--- a/crates/benilla-app/src/post/sun_shafts.rs
+++ b/crates/benilla-app/src/post/sun_shafts.rs
@@ -165,6 +165,10 @@ const MASK_FORMAT: TextureFormat = TextureFormat::R8Unorm;
 #[derive(Component)]
 struct ShaftMask(CachedTexture);
 
+/// MONKEY (polish): the view's uniform, created once and rewritten in prepare, like bloom's.
+#[derive(Component)]
+struct ShaftUniform(Buffer);
+
 #[derive(Component)]
 struct ViewShaftPipeline(CachedRenderPipelineId);
 
@@ -277,17 +281,31 @@ fn prepare_pipelines(
     }
 }
 
-/// The half-size mask target, rounded up so an odd edge column still has a texel.
+/// The half-size mask target, rounded up so an odd edge column still has a texel, and the
+/// view's uniform, written in place.
 fn prepare_masks(
     mut commands: Commands,
     mut textures: ResMut<TextureCache>,
     device: Res<RenderDevice>,
-    views: Query<(Entity, &ExtractedCamera), With<ShaftView>>,
+    queue: Res<RenderQueue>,
+    views: Query<(Entity, &ExtractedCamera, &ShaftView, Option<&ShaftUniform>)>,
 ) {
-    for (entity, camera) in &views {
+    for (entity, camera, shaft, uniform) in &views {
         let Some(size) = camera.physical_viewport_size else {
             continue;
         };
+        let rows = [shaft.sun.to_array(), shaft.color.to_array()];
+        match uniform {
+            Some(uniform) => queue.write_buffer(&uniform.0, 0, bytemuck::cast_slice(&rows)),
+            None => {
+                let buffer = device.create_buffer_with_data(&BufferInitDescriptor {
+                    label: Some("post_sun_shafts_uniform"),
+                    contents: bytemuck::cast_slice(&rows),
+                    usage: BufferUsages::UNIFORM | BufferUsages::COPY_DST,
+                });
+                commands.entity(entity).insert(ShaftUniform(buffer));
+            }
+        }
         let mask = textures.get(
             &device,
             TextureDescriptor {
@@ -316,16 +334,18 @@ impl ViewNode for ShaftNode {
     type ViewQuery = (
         &'static ViewTarget,
         &'static ViewDepthTexture,
+        // Gates the pass: the mask and uniform outlive a disabled frame.
         &'static ShaftView,
         &'static ViewShaftPipeline,
         &'static ShaftMask,
+        &'static ShaftUniform,
     );
 
     fn run<'w>(
         &self,
         _graph: &mut RenderGraphContext,
         context: &mut RenderContext<'w>,
-        (target, depth, shaft, id, mask): QueryItem<'w, '_, Self::ViewQuery>,
+        (target, depth, _, id, mask, uniform): QueryItem<'w, '_, Self::ViewQuery>,
         world: &'w World,
     ) -> Result<(), NodeRunError> {
         let cache = world.resource::<PipelineCache>();
@@ -350,8 +370,6 @@ impl ViewNode for ShaftNode {
             &cache.get_bind_group_layout(&settings.mask_layouts[multisampled]),
             &BindGroupEntries::single(depth.view()),
         );
-        let mut uniform = UniformBuffer::from(*shaft);
-        uniform.write_buffer(device, world.resource::<RenderQueue>());
         let out = target.post_process_write();
         let bind = device.create_bind_group(
             "post_sun_shafts",
@@ -360,7 +378,7 @@ impl ViewNode for ShaftNode {
                 out.source,
                 &settings.sampler,
                 &mask.0.default_view,
-                uniform.binding().unwrap(),
+                uniform.0.as_entire_binding(),
             )),
         );
         let diagnostics = context.diagnostic_recorder();

From 7d2f55a0147c50d02b8d903bf7cd525a43726162 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Fri, 25 Sep 2026 18:38:38 +0000
Subject: [PATCH 3/5] MONKEY (polish): crossfade colour grading across zone
 borders

The grade snapped to the new zone's LUT pair on the frame `CurrentArea`
changed. `GradeBlend` now holds the current and previous pairs and fades
between them over GRADE_FADE_S = 2 s: the view binds a second day/night
pair (bindings 6, 7; identity when there is none) and `control.z` is the
current pair's weight, `control.w` the previous pair's strength. The shader
runs the previous pair through the same day/night blend at its own
strength and mixes, only while z < 1. A zone with no row fades to identity
(strength 0) instead of cutting; stepping back across the border mid-fade
reverses the fade from the image on screen. The first graded zone after
login, and a disabled grade, snap. Missing LUTs still bind the identity cube.

Verified: cargo test -p benilla-app --lib post:: (4 passed, including the
new zone_change_crossfades_and_a_quick_return_reverses); grading.wgsl
composes and validates under naga_oil 0.20 / naga 27 offline. No GPU or
WoW data here; built on Linux only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y6yUWgPosSakvL23gtgvpH
---
 crates/benilla-app/src/post/grading.rs   | 142 ++++++++++++++++++++++-
 crates/benilla-app/src/post/grading.wgsl |  14 ++-
 2 files changed, 151 insertions(+), 5 deletions(-)

diff --git a/crates/benilla-app/src/post/grading.rs b/crates/benilla-app/src/post/grading.rs
index b3fd694d..7d2b28d6 100644
--- a/crates/benilla-app/src/post/grading.rs
+++ b/crates/benilla-app/src/post/grading.rs
@@ -43,10 +43,66 @@ struct GradeCatalog {
 struct GradeView {
     day: Handle<Image>,
     night: Handle<Image>,
-    // x = authored strength, y = night blend.
+    /// MONKEY (polish): the pair being faded out (identity once the fade is done).
+    prev_day: Handle<Image>,
+    prev_night: Handle<Image>,
+    // x = authored strength, y = night blend, z = crossfade weight of the current pair (1 = done),
+    // w = the previous pair's strength.
     control: Vec4,
 }
 
+/// MONKEY (polish): the zone crossfade's length; a border crossing no longer snaps the grade.
+const GRADE_FADE_S: f32 = 2.0;
+
+/// One zone's LUT pair at its authored strength; a zone with no row grades at strength 0.
+#[derive(Clone, PartialEq, Debug)]
+struct GradePair {
+    day: Handle<Image>,
+    night: Handle<Image>,
+    strength: f32,
+}
+
+/// MONKEY (polish): the crossfade from the previous zone's pair to the current one.
+#[derive(Resource, Default)]
+struct GradeBlend {
+    current: Option<GradePair>,
+    previous: Option<GradePair>,
+    /// The current pair's weight, 0..1.
+    fade: f32,
+    primed: bool,
+}
+
+impl GradeBlend {
+    /// Steps toward `target`. The first graded zone after login (or re-enabling) and a disabled
+    /// grade snap; a return to the
+    /// pair still fading out reverses the fade instead of restarting it.
+    fn step(&mut self, target: Option<GradePair>, dt: f32, snap: bool) {
+        if snap || !self.primed {
+            *self = Self {
+                primed: !snap && target.is_some(),
+                current: target,
+                previous: None,
+                fade: 1.0,
+            };
+            return;
+        }
+        if target != self.current {
+            if self.fade < 1.0 && target == self.previous {
+                std::mem::swap(&mut self.current, &mut self.previous);
+                self.fade = 1.0 - self.fade;
+            } else {
+                self.previous = self.current.take();
+                self.current = target;
+                self.fade = 0.0;
+            }
+        }
+        self.fade = (self.fade + dt / GRADE_FADE_S).min(1.0);
+        if self.fade >= 1.0 {
+            self.previous = None;
+        }
+    }
+}
+
 #[derive(Clone, Copy, ShaderType)]
 struct GradeUniform {
     control: Vec4,
@@ -55,6 +111,7 @@ struct GradeUniform {
 impl Plugin for GradingPlugin {
     fn build(&self, app: &mut App) {
         app.add_plugins(ExtractComponentPlugin::<GradeView>::default())
+            .init_resource::<GradeBlend>()
             .add_systems(Last, update_views);
         if !app.is_plugin_added::<AssetPlugin>() {
             return;
@@ -219,6 +276,7 @@ fn volume_image(cube: Vec<u8>) -> Image {
     )
 }
 
+#[allow(clippy::too_many_arguments)]
 fn update_views(
     mut commands: Commands,
     video: Res<VideoConfig>,
@@ -226,6 +284,8 @@ fn update_views(
     area: Res<CurrentArea>,
     areas: Option<Res<crate::area::AreaTableRes>>,
     time: Res<WorldTime>,
+    clock: Res<Time>,
+    mut blend: ResMut<GradeBlend>,
     cameras: Query<(Entity, &Camera, Option<&GradeView>), With<WorldCamera>>,
 ) {
     let selected = catalog.as_ref().and_then(|catalog| {
@@ -238,7 +298,7 @@ fn update_views(
             .rows
             .for_area(leaf)
             .or_else(|| catalog.rows.for_area(zone))?;
-        (row.strength > 0.0).then(|| GradeView {
+        (row.strength > 0.0).then(|| GradePair {
             day: catalog
                 .luts
                 .get(&row.day_lut)
@@ -249,7 +309,33 @@ fn update_views(
                 .get(&row.night_lut)
                 .unwrap_or(&catalog.identity)
                 .clone(),
-            control: Vec4::new(row.strength, night_weight(time.minute_f), 0.0, 0.0),
+            strength: row.strength,
+        })
+    });
+    blend.step(selected, clock.delta_secs(), !video.color_grading);
+    let selected = catalog.as_ref().and_then(|catalog| {
+        if blend.current.is_none() && blend.previous.is_none() {
+            return None;
+        }
+        let pair = |p: &Option<GradePair>| {
+            p.as_ref().map_or_else(
+                || (catalog.identity.clone(), catalog.identity.clone(), 0.0),
+                |p| (p.day.clone(), p.night.clone(), p.strength),
+            )
+        };
+        let (day, night, strength) = pair(&blend.current);
+        let (prev_day, prev_night, prev_strength) = pair(&blend.previous);
+        Some(GradeView {
+            day,
+            night,
+            prev_day,
+            prev_night,
+            control: Vec4::new(
+                strength,
+                night_weight(time.minute_f),
+                blend.fade,
+                prev_strength,
+            ),
         })
     });
     for (entity, camera, old) in &cameras {
@@ -311,6 +397,9 @@ fn init_pipeline(
                 texture_3d(TextureSampleType::Float { filterable: true }),
                 sampler(SamplerBindingType::Filtering),
                 uniform_buffer::<GradeUniform>(false),
+                // MONKEY (polish): the previous zone's pair, crossfaded out.
+                texture_3d(TextureSampleType::Float { filterable: true }),
+                texture_3d(TextureSampleType::Float { filterable: true }),
             ),
         ),
     );
@@ -397,7 +486,12 @@ impl ViewNode for GradeNode {
             return Ok(());
         };
         let images = world.resource::<RenderAssets<GpuImage>>();
-        let (Some(day), Some(night)) = (images.get(&grade.day), images.get(&grade.night)) else {
+        let (Some(day), Some(night), Some(prev_day), Some(prev_night)) = (
+            images.get(&grade.day),
+            images.get(&grade.night),
+            images.get(&grade.prev_day),
+            images.get(&grade.prev_night),
+        ) else {
             return Ok(());
         };
         let device = context.render_device();
@@ -412,6 +506,8 @@ impl ViewNode for GradeNode {
                 &night.texture_view,
                 &settings.lut_sampler,
                 uniform.0.as_entire_binding(),
+                &prev_day.texture_view,
+                &prev_night.texture_view,
             )),
         );
         let diagnostics = context.diagnostic_recorder();
@@ -452,6 +548,44 @@ mod tests {
         assert_eq!(at(0, 0, 31), &[0, 0, 255, 255]);
     }
 
+    #[test]
+    fn zone_change_crossfades_and_a_quick_return_reverses() {
+        let pair = |id: u128, strength| GradePair {
+            day: Handle::Uuid(bevy::asset::uuid::Uuid::from_u128(id), default()),
+            night: Handle::Uuid(bevy::asset::uuid::Uuid::from_u128(id + 1), default()),
+            strength,
+        };
+        let (a, b) = (pair(10, 1.0), pair(20, 0.5));
+        let mut blend = GradeBlend::default();
+        // No zone yet, then the first graded zone snaps in.
+        blend.step(None, 0.016, false);
+        blend.step(Some(a.clone()), 0.016, false);
+        assert_eq!((blend.current.clone(), blend.fade), (Some(a.clone()), 1.0));
+        // A border crossing fades over GRADE_FADE_S.
+        blend.step(Some(b.clone()), 0.5, false);
+        assert_eq!(blend.previous, Some(a.clone()));
+        assert!((blend.fade - 0.25).abs() < 1e-6);
+        // Stepping back mid-fade reverses from the same image.
+        blend.step(Some(a.clone()), 0.0, false);
+        assert_eq!(
+            (blend.current.clone(), blend.previous.clone()),
+            (Some(a.clone()), Some(b))
+        );
+        assert!((blend.fade - 0.75).abs() < 1e-6);
+        blend.step(Some(a.clone()), 1.0, false);
+        assert_eq!((blend.fade, blend.previous.is_none()), (1.0, true));
+        // An ungraded zone fades out to identity rather than cutting.
+        blend.step(None, 1.0, false);
+        assert_eq!(
+            (blend.current.is_none(), blend.previous.clone()),
+            (true, Some(a))
+        );
+        assert!((blend.fade - 0.5).abs() < 1e-6);
+        // Disabling snaps.
+        blend.step(None, 0.0, true);
+        assert!(blend.previous.is_none() && blend.current.is_none());
+    }
+
     #[test]
     fn day_night_blend_is_continuous() {
         assert_eq!(night_weight(0.0), 1.0);
diff --git a/crates/benilla-app/src/post/grading.wgsl b/crates/benilla-app/src/post/grading.wgsl
index fefdc9da..103ef98b 100644
--- a/crates/benilla-app/src/post/grading.wgsl
+++ b/crates/benilla-app/src/post/grading.wgsl
@@ -15,6 +15,9 @@ struct Grade {
     control: vec4<f32>,
 };
 @group(0) @binding(5) var<uniform> grade: Grade;
+// MONKEY (polish): the previous zone's pair, crossfaded out over `grade.control.z`.
+@group(0) @binding(6) var prev_day_lut: texture_3d<f32>;
+@group(0) @binding(7) var prev_night_lut: texture_3d<f32>;
 
 fn grade_one(lut: texture_3d<f32>, color: vec3<f32>) -> vec3<f32> {
     // Half a texel in from each face, with 31 intervals across the 32-texel cube.
@@ -29,7 +32,16 @@ fn fragment(in: FullscreenVertexOutput) -> @location(0) vec4<f32> {
     let day = grade_one(day_lut, base);
     let night = grade_one(night_lut, base);
     let authored = mix(day, night, grade.control.y);
-    let graded = mix(base, authored, grade.control.x);
+    var graded = mix(base, authored, grade.control.x);
+    // MONKEY (polish): mid-crossfade, the previous zone's pair through the same day/night blend
+    // at its own strength; a missing zone or LUT is the identity cube, so it fades to ungraded.
+    if (grade.control.z < 1.0) {
+        let prev_day = grade_one(prev_day_lut, base);
+        let prev_night = grade_one(prev_night_lut, base);
+        let prev_authored = mix(prev_day, prev_night, grade.control.y);
+        let prev_graded = mix(base, prev_authored, grade.control.w);
+        graded = mix(prev_graded, graded, grade.control.z);
+    }
     // Bloom already consumed the HDR signal; preserve its remaining excess for the FFX clamp.
     return vec4(graded + max(source.rgb - vec3(1.0), vec3(0.0)), source.a);
 }

From 0412a7f362552f9ce72bc519f6d0bfc4aff27510 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Fri, 25 Sep 2026 19:02:03 +0000
Subject: [PATCH 4/5] MONKEY (polish): the sky clock leaves the sky and cloud
 materials

At skyQuality >= 1 the dome and cloud materials were rewritten every frame
only because `fx.y` carried the sky clock, so each frame re-prepared both
material bind groups. The clock now lives in `SkyClockBuffer`, a 16-byte
STORAGE | COPY_DST buffer created beside the shared light buffer
(AssetSet::Open) and bound read-only by both materials (sky binding 101,
cloud binding 103), the same shape as the materials' shared `light_buf`.
The render world writes `SkyClock::secs` into it in PrepareResources each
frame. `fx.y` is now reserved (0), so `write_gated` only touches a material
when a real input moves. The clock is still our own f64 accumulator with
the seamless 86400 s wrap (Bevy's globals.time wraps hourly); capture mode
still freezes it. A headless build with no RenderDevice gets no buffer and
spawns no sky or cloud dome.

Verified: cargo check -p benilla; cargo test -p benilla-world --lib (643
passed, 8 ignored); sky.wgsl (Classic and SKY_FX) and cloud.wgsl (plain and
SKY_FX_CLOUDS + VERTEX_COLORS) compose and validate under naga_oil 0.20 /
naga 27 offline against stubbed bevy_pbr imports and the real fog_hook,
monkey_frame and sky_fx modules. No GPU or WoW data here; built on Linux only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y6yUWgPosSakvL23gtgvpH
---
 crates/benilla-world/src/clouds/layer.rs    | 19 ++++--
 crates/benilla-world/src/shaders/cloud.wgsl |  8 ++-
 crates/benilla-world/src/shaders/sky.wgsl   |  6 +-
 crates/benilla-world/src/sky.rs             | 22 ++++---
 crates/benilla-world/src/sky_fx.rs          | 67 ++++++++++++++++++++-
 5 files changed, 104 insertions(+), 18 deletions(-)

diff --git a/crates/benilla-world/src/clouds/layer.rs b/crates/benilla-world/src/clouds/layer.rs
index fc495faf..cfffa46e 100644
--- a/crates/benilla-world/src/clouds/layer.rs
+++ b/crates/benilla-world/src/clouds/layer.rs
@@ -12,7 +12,7 @@ use bevy::pbr::{
 };
 use bevy::prelude::*;
 use bevy::render::render_resource::{
-    AsBindGroup, Extent3d, RenderPipelineDescriptor, SpecializedMeshPipelineError,
+    AsBindGroup, Buffer, Extent3d, RenderPipelineDescriptor, SpecializedMeshPipelineError,
     TextureDimension, TextureFormat,
 };
 use bevy::shader::ShaderRef;
@@ -35,13 +35,18 @@ pub struct CloudExt {
     #[texture(100)]
     #[sampler(101)]
     pub(crate) texels: Handle<Image>,
-    /// MONKEY (sky): `x` the sky tier (detail at 2), `y` the sky clock, `zw` the unit direction
-    /// toward the glow body on the sheet ([`update_cloud_fx`]).
+    /// MONKEY (sky): `x` the sky tier (detail at 2), `y` reserved (0; the clock is
+    /// [`Self::clock`]), `zw` the unit direction toward the glow body on the sheet
+    /// ([`update_cloud_fx`]).
     #[uniform(102)]
     pub(crate) fx: Vec4,
     /// MONKEY (sky): the Light.dbc cloud sun colour (`rgb`, gamma) and the march strength (`w`).
     #[uniform(102)]
     pub(crate) lit: Vec4,
+    /// MONKEY (polish): the shared sky clock ([`crate::sky_fx::SkyClockBuffer`]), outside the
+    /// uniform so the drift never rewrites the material.
+    #[storage(103, read_only, buffer)]
+    pub(crate) clock: Buffer,
 }
 
 /// MONKEY (sky): the pipeline key: High compiles `SKY_FX_CLOUDS` into `cloud.wgsl`; below High
@@ -159,7 +164,11 @@ pub(super) fn setup_cloud_layer(
     mut meshes: ResMut<Assets<Mesh>>,
     mut materials: ResMut<Assets<CloudMaterial>>,
     mut images: ResMut<Assets<Image>>,
+    clock: Option<Res<crate::sky_fx::SkyClockBuffer>>,
 ) {
+    let Some(clock) = clock else {
+        return;
+    };
     let image = images.add(Image::new(
         Extent3d {
             width: COLS as u32,
@@ -186,6 +195,7 @@ pub(super) fn setup_cloud_layer(
             // MONKEY (sky): Classic until `update_cloud_fx` reads the tier.
             fx: Vec4::ZERO,
             lit: Vec4::ZERO,
+            clock: clock.0.clone(),
         },
     });
     commands.spawn((
@@ -251,7 +261,6 @@ pub(super) fn follow_cloud_dome(
 pub(super) fn update_cloud_fx(
     light: Res<crate::lighting::WowLighting>,
     quality: Res<crate::sky_fx::SkyQuality>,
-    clock: Res<crate::sky_fx::SkyClock>,
     layer: Option<Res<CloudLayer>>,
     mut materials: ResMut<Assets<CloudMaterial>>,
 ) {
@@ -268,7 +277,7 @@ pub(super) fn update_cloud_fx(
         let q = |v: f32| benilla_assets::quantize(v, 4096.0);
         let sun = light.cloud_colors[0];
         (
-            Vec4::new(f32::from(quality.0), clock.secs, q(d.x), q(d.y)),
+            Vec4::new(f32::from(quality.0), 0.0, q(d.x), q(d.y)),
             Vec4::new(q(sun[0]), q(sun[1]), q(sun[2]), q(strength.clamp(0.0, 1.0))),
         )
     } else {
diff --git a/crates/benilla-world/src/shaders/cloud.wgsl b/crates/benilla-world/src/shaders/cloud.wgsl
index d2413618..36167bea 100644
--- a/crates/benilla-world/src/shaders/cloud.wgsl
+++ b/crates/benilla-world/src/shaders/cloud.wgsl
@@ -14,7 +14,7 @@
 @group(#{MATERIAL_BIND_GROUP}) @binding(100) var cloud_tex: texture_2d<f32>;
 @group(#{MATERIAL_BIND_GROUP}) @binding(101) var cloud_samp: sampler;
 
-// MONKEY (sky): `fx.x` the sky tier, `.y` the sky clock (s), `.zw` the unit direction toward the
+// MONKEY (sky): `fx.x` the sky tier, `.y` reserved, `.zw` the unit direction toward the
 // glow body on the sheet (u = world x, v = world z); `lit.rgb` the Light.dbc cloud sun colour
 // (gamma), `.w` the march strength (the body's flatness × the glow envelope).
 struct CloudFx {
@@ -22,6 +22,8 @@ struct CloudFx {
     lit: vec4<f32>,
 };
 @group(#{MATERIAL_BIND_GROUP}) @binding(102) var<uniform> cfx: CloudFx;
+// MONKEY (polish): x = the sky clock (s, wrapped at a day), written per frame outside the material.
+@group(#{MATERIAL_BIND_GROUP}) @binding(103) var<storage, read> sky_clock: vec4<f32>;
 
 #ifdef SKY_FX_CLOUDS
 // MONKEY (sky): coverage with detail at a sheet point; `oct` detail octaves.
@@ -30,7 +32,7 @@ fn detailed_cover(uv: vec2<f32>, oct: i32) -> f32 {
     if (a <= 0.0) {
         return 0.0;
     }
-    return sky_fx::cloud_erode(a, sky_fx::cloud_detail(uv, cfx.fx.y, oct));
+    return sky_fx::cloud_erode(a, sky_fx::cloud_detail(uv, sky_clock.x, oct));
 }
 #endif
 
@@ -43,7 +45,7 @@ fn fragment(in: VertexOutput) -> @location(0) vec4<f32> {
     // kernel's texels as they are.
 #ifdef SKY_FX_CLOUDS
     if (cfx.fx.x >= 1.5) {
-        let n = sky_fx::cloud_detail(in.uv, cfx.fx.y, 4);
+        let n = sky_fx::cloud_detail(in.uv, sky_clock.x, 4);
         a = sky_fx::cloud_erode(a, n);
         // A painterly tone inside the mass from the same detail.
         rgb = rgb * (0.93 + 0.14 * n);
diff --git a/crates/benilla-world/src/shaders/sky.wgsl b/crates/benilla-world/src/shaders/sky.wgsl
index bf732117..5301a4a2 100644
--- a/crates/benilla-world/src/shaders/sky.wgsl
+++ b/crates/benilla-world/src/shaders/sky.wgsl
@@ -21,7 +21,7 @@ struct SkyColors {
     sky4: vec4<f32>, // 1.8°
     fog: vec4<f32>,  // horizon (0°) and below: LightIntBand row 7
     warp: vec4<f32>, // x = dawn/dusk warp strength S (0 = off), y = sun azimuth (rad), zw reserved
-    // MONKEY (sky): x = skyQuality (0 Classic), y = sky clock (s), z = night-sky alpha, w = glow
+    // MONKEY (sky): x = skyQuality (0 Classic), y = reserved, z = night-sky alpha, w = glow
     // strength.
     fx: vec4<f32>,
     // MONKEY (sky): xyz = camera to the visible sun, w unused.
@@ -34,6 +34,8 @@ struct SkyColors {
     mf_fog_d: vec4<f32>,
 };
 @group(#{MATERIAL_BIND_GROUP}) @binding(100) var<uniform> sky: SkyColors;
+// MONKEY (polish): x = the sky clock (s, wrapped at a day), written per frame outside the material.
+@group(#{MATERIAL_BIND_GROUP}) @binding(101) var<storage, read> sky_clock: vec4<f32>;
 
 // Glow `g` for a sun-relative azimuth phase, sampled linearly like `0x6d0f50` from the reference's
 // six-keyframe wrap-around table at `0xce9af8` (written by `0x6ce210`); 0.125 is the sun bearing.
@@ -137,7 +139,7 @@ fn fragment(in: VertexOutput) -> @location(0) vec4<f32> {
         let px = length(fwidth(dir)) * sky_fx::STAR_GRID * 0.5;
         if (night > 0.0 && dir.y > 0.0) {
             let band = sky_fx::galaxy_band(dir);
-            let stars = sky_fx::star_field(dir, sky.fx.y, max(px, 1e-4), band)
+            let stars = sky_fx::star_field(dir, sky_clock.x, max(px, 1e-4), band)
                 * smoothstep(0.0, 0.2, dir.y);
             let milky = sky_fx::milky_way(dir, band) * 0.05 * smoothstep(0.03, 0.35, dir.y);
             lin += (stars * 0.85 + milky) * night;
diff --git a/crates/benilla-world/src/sky.rs b/crates/benilla-world/src/sky.rs
index 90552c67..a1bcec7e 100644
--- a/crates/benilla-world/src/sky.rs
+++ b/crates/benilla-world/src/sky.rs
@@ -17,7 +17,7 @@ use bevy::pbr::{
 };
 use bevy::prelude::*;
 use bevy::render::render_resource::{
-    AsBindGroup, RenderPipelineDescriptor, SpecializedMeshPipelineError,
+    AsBindGroup, Buffer, RenderPipelineDescriptor, SpecializedMeshPipelineError,
 };
 use bevy::shader::ShaderRef;
 
@@ -57,7 +57,8 @@ pub struct SkyExt {
     #[uniform(100)]
     pub(crate) warp: Vec4,
     /// MONKEY (sky): `x` the [`crate::sky_fx::SkyQuality`] tier (0 = Classic, the fields below
-    /// unused), `y` the sky clock in seconds, `z` the night-sky alpha, `w` the sun-glow strength.
+    /// unused), `y` reserved (0; the clock is [`Self::clock`]), `z` the night-sky alpha, `w` the
+    /// sun-glow strength.
     #[uniform(100)]
     pub(crate) fx: Vec4,
     /// MONKEY (sky): camera to the visible sun (`xyz`).
@@ -73,6 +74,10 @@ pub struct SkyExt {
     pub(crate) mf_fog_c: Vec4,
     #[uniform(100)]
     pub(crate) mf_fog_d: Vec4,
+    /// MONKEY (polish): the shared sky clock ([`crate::sky_fx::SkyClockBuffer`]), outside the
+    /// uniform so the clock never rewrites the material.
+    #[storage(101, read_only, buffer)]
+    pub(crate) clock: Buffer,
 }
 
 /// MONKEY (sky): the pipeline key: Enhanced/High compile `SKY_FX` into `sky.wgsl`; Classic keeps
@@ -206,7 +211,11 @@ fn setup_sky(
     mut commands: Commands,
     mut meshes: ResMut<Assets<Mesh>>,
     mut materials: ResMut<Assets<SkyMaterial>>,
+    clock: Option<Res<crate::sky_fx::SkyClockBuffer>>,
 ) {
+    let Some(clock) = clock else {
+        return;
+    };
     let mesh = meshes.add(dome_mesh());
     // Seeded with the neutral-day fallback; `update_sky_colors` overwrites from Light.dbc each frame.
     let material = materials.add(SkyMaterial {
@@ -231,6 +240,7 @@ fn setup_sky(
             mf_fog_b: Vec4::ZERO,
             mf_fog_c: Vec4::ZERO,
             mf_fog_d: Vec4::ZERO,
+            clock: clock.0.clone(),
         },
     });
     commands.spawn((
@@ -293,9 +303,8 @@ fn apply_sky_visibility(
 #[allow(clippy::too_many_arguments)]
 fn update_sky_colors(
     light: Res<WowLighting>,
-    // MONKEY (sky): the tier, the sky clock and the cloud cover over the sun.
+    // MONKEY (sky): the tier and the cloud cover over the sun.
     quality: Res<crate::sky_fx::SkyQuality>,
-    clock: Res<crate::sky_fx::SkyClock>,
     clouds: Res<crate::clouds::CloudCoverage>,
     // MONKEY (fog): the Modern fog rows for the shared horizon colour.
     monkey: Res<crate::lighting::MonkeyFrame>,
@@ -326,7 +335,7 @@ fn update_sky_colors(
         0.0,
     );
     // MONKEY (sky): the Enhanced/High inputs; all zero at Classic so the write gate stays quiet.
-    let (fx, sun_v, glow) = sky_fx_inputs(&light, *quality, clock.secs, &clouds);
+    let (fx, sun_v, glow) = sky_fx_inputs(&light, *quality, &clouds);
     // MONKEY (fog): rows B-D as packed for the light buffer; clock lanes are not read here.
     let rows = monkey.pack(0.0, 0.0);
     let mf_fog = [
@@ -374,7 +383,6 @@ fn update_sky_colors(
 fn sky_fx_inputs(
     light: &WowLighting,
     quality: crate::sky_fx::SkyQuality,
-    secs: f32,
     clouds: &crate::clouds::CloudCoverage,
 ) -> (Vec4, Vec4, Vec4) {
     if !quality.enhanced() {
@@ -393,7 +401,7 @@ fn sky_fx_inputs(
     // Quantized like every other lane here, so an idle frame writes nothing.
     let q = |v: f32| benilla_assets::quantize(v, 4096.0);
     (
-        Vec4::new(f32::from(quality.0), secs, q(night), q(strength)),
+        Vec4::new(f32::from(quality.0), 0.0, q(night), q(strength)),
         Vec4::new(q(sun.x), q(sun.y), q(sun.z), 0.0),
         Vec4::new(q(glow.x), q(glow.y), q(glow.z), 0.0),
     )
diff --git a/crates/benilla-world/src/sky_fx.rs b/crates/benilla-world/src/sky_fx.rs
index e8462579..2c6accc4 100644
--- a/crates/benilla-world/src/sky_fx.rs
+++ b/crates/benilla-world/src/sky_fx.rs
@@ -9,6 +9,12 @@
 //! Ported from WarcraftXL (https://github.com/WarcraftXL) by iThorgrim — module wxl-retail-clouds, Clouds.cpp, Clouds.hpp.
 
 use bevy::prelude::*;
+use bevy::render::{
+    extract_resource::{ExtractResource, ExtractResourcePlugin},
+    render_resource::{Buffer, BufferDescriptor, BufferUsages},
+    renderer::{RenderDevice, RenderQueue},
+    Render, RenderApp, RenderSystems,
+};
 
 /// The sky tier: 0 Classic, 1 Enhanced, 2 High. Written by the app from the `skyQuality` cvar;
 /// `$WOW_SKY_QUALITY` pins it for the session (captures).
@@ -53,6 +59,25 @@ pub struct SkyClock {
 /// The sky clock's wrap in seconds; `sky_fx.wgsl`'s `SKY_WRAP` must match.
 pub const SKY_CLOCK_WRAP_S: f64 = 86_400.0;
 
+/// MONKEY (polish): the sky clock on the GPU, one `vec4` (`x` = [`SkyClock::secs`]) that the sky
+/// and cloud materials bind read-only. It is rewritten in place each frame in the render world,
+/// so the materials re-prepare only when a real input moves, not to advance the clock. Bevy's
+/// `globals.time` is not used: it wraps hourly, and the twinkle rates are whole cycles per day.
+#[derive(Resource, Clone, ExtractResource)]
+pub struct SkyClockBuffer(pub Buffer);
+
+/// The clock value the render world uploads.
+#[derive(Resource, Clone, Copy)]
+struct SkyClockSecs(f32);
+
+impl ExtractResource for SkyClockSecs {
+    type Source = SkyClock;
+
+    fn extract_resource(source: &SkyClock) -> Self {
+        Self(source.secs)
+    }
+}
+
 /// How strongly the glow shows: a broad halo at `GLOW_GAIN` × the sun colour at the sun itself.
 pub(crate) const GLOW_GAIN: f32 = 0.22;
 
@@ -91,10 +116,50 @@ impl Plugin for SkyFxPlugin {
         });
         let quality = SkyQuality(SkyQuality::env_override().unwrap_or(0));
         app.insert_resource(quality)
-            .add_systems(Update, tick_sky_clock);
+            .add_systems(Update, tick_sky_clock)
+            // Beside the shared light buffer, ahead of the sky and cloud materials built after it.
+            .add_systems(
+                Startup,
+                init_sky_clock_buffer.in_set(benilla_assets::AssetSet::Open),
+            )
+            .add_plugins((
+                ExtractResourcePlugin::<SkyClockBuffer>::default(),
+                ExtractResourcePlugin::<SkyClockSecs>::default(),
+            ));
+        if let Some(render_app) = app.get_sub_app_mut(RenderApp) {
+            render_app.add_systems(
+                Render,
+                upload_sky_clock.in_set(RenderSystems::PrepareResources),
+            );
+        }
     }
 }
 
+/// A headless build has no device, and then no sky or cloud dome either.
+fn init_sky_clock_buffer(mut commands: Commands, device: Option<Res<RenderDevice>>) {
+    let Some(device) = device else {
+        return;
+    };
+    commands.insert_resource(SkyClockBuffer(device.create_buffer(&BufferDescriptor {
+        label: Some("sky_clock"),
+        size: 16,
+        usage: BufferUsages::STORAGE | BufferUsages::COPY_DST,
+        mapped_at_creation: false,
+    })));
+}
+
+/// Render-world: the clock into the shared buffer before any sky draw reads it.
+fn upload_sky_clock(
+    queue: Res<RenderQueue>,
+    buffer: Option<Res<SkyClockBuffer>>,
+    secs: Option<Res<SkyClockSecs>>,
+) {
+    let (Some(buffer), Some(secs)) = (buffer, secs) else {
+        return;
+    };
+    queue.write_buffer(&buffer.0, 0, bytemuck::cast_slice(&[secs.0, 0.0, 0.0, 0.0]));
+}
+
 fn tick_sky_clock(time: Res<Time>, mut clock: ResMut<SkyClock>) {
     if let Some(t) = clock.frozen {
         if clock.secs != t {

From 111742f83e894941fd769dc3f1a623bf961f0ede Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Fri, 25 Sep 2026 19:03:50 +0000
Subject: [PATCH 5/5] MONKEY (polish): the High cloud detail loops over the sky
 clock's day wrap

The detail drift was `(0.011, 0.0043) * t`, so the field jumped to a new
pattern when the sky clock wrapped at 86400 s. `value_noise_plane` now takes
a lattice period and hashes cell indices modulo it; the detail plane tiles
every CLOUD_TILE = 64 cells (octave o every 64 << o, the half-rate domain
warp every 32), and the drift is whole tiles per wrap, CLOUD_DRIFT_TILES =
(15, 6), i.e. 0.01111 and 0.00444 cells/s (+1% and +3% over WarcraftXL's
rates). The shader takes the drift as `fract(tiles * t / SKY_WRAP) * tile`,
so the field at the wrap equals the field at 0, and the drift term stays
below 64 instead of growing to ~950 in f32. One tile spans 1.78 of the
sheet's UV, wider than the dome shows, so no spatial repeat is visible.

A test reads `sky_fx.wgsl` and pins SKY_WRAP to SKY_CLOCK_WRAP_S, an even
tile, whole tiles of drift and the rates within 5% of the originals.

Verified: cargo test -p benilla-world --lib (644 passed, 8 ignored);
cloud.wgsl (SKY_FX_CLOUDS) and sky.wgsl (SKY_FX) compose and validate under
naga_oil 0.20 / naga 27 offline. No GPU or WoW data here; built on Linux only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y6yUWgPosSakvL23gtgvpH
---
 crates/benilla-world/src/shaders/sky_fx.wgsl | 35 +++++++++++++-----
 crates/benilla-world/src/sky_fx.rs           | 37 ++++++++++++++++++--
 2 files changed, 62 insertions(+), 10 deletions(-)

diff --git a/crates/benilla-world/src/shaders/sky_fx.wgsl b/crates/benilla-world/src/shaders/sky_fx.wgsl
index 251fcb04..e9f8f16c 100644
--- a/crates/benilla-world/src/shaders/sky_fx.wgsl
+++ b/crates/benilla-world/src/shaders/sky_fx.wgsl
@@ -54,14 +54,18 @@ fn hash_cell(p: vec2<i32>, salt: u32) -> f32 {
 
 // Value noise with the quintic fade: the cubic's second-derivative seam shows as square plateaus
 // on a slowly drifting sky (the WarcraftXL note).
-fn value_noise_plane(p: vec2<f32>, salt: u32) -> f32 {
+// MONKEY (polish): the lattice repeats every `period` cells, so a shift by `period` is no change.
+fn value_noise_plane(p: vec2<f32>, salt: u32, period: i32) -> f32 {
     let i = vec2<i32>(floor(p));
     let f = fract(p);
     let u = f * f * f * (f * (f * 6.0 - 15.0) + 10.0);
-    let a = hash_cell(i, salt);
-    let b = hash_cell(i + vec2<i32>(1, 0), salt);
-    let c = hash_cell(i + vec2<i32>(0, 1), salt);
-    let d = hash_cell(i + vec2<i32>(1, 1), salt);
+    let per = vec2<i32>(period);
+    let lo = ((i % per) + per) % per;
+    let hi = (lo + vec2<i32>(1)) % per;
+    let a = hash_cell(lo, salt);
+    let b = hash_cell(vec2<i32>(hi.x, lo.y), salt);
+    let c = hash_cell(vec2<i32>(lo.x, hi.y), salt);
+    let d = hash_cell(hi, salt);
     return mix(mix(a, b, u.x), mix(c, d, u.x), u.y);
 }
 
@@ -244,6 +248,15 @@ fn milky_way(dir: vec3<f32>, band: f32) -> vec3<f32> {
 
 // ---- S3: cloud detail -------------------------------------------------------------------------
 
+// MONKEY (polish): the detail field loops over the sky clock's wrap. It tiles every `CLOUD_TILE`
+// cells of the detail plane (octave o every `CLOUD_TILE << o` of its own lattice; the half-rate
+// warp every `CLOUD_TILE / 2`), and the drift covers whole tiles per wrap, so the field at
+// `SKY_WRAP` is the field at 0. One tile is 64 / 36 of the sheet, wider than the dome's UV.
+const CLOUD_TILE: i32 = 64;
+// Tiles of drift per wrap on u and v: 0.01111 and 0.00444 cells/s, WarcraftXL's 0.011 and 0.0043
+// rounded to whole tiles.
+const CLOUD_DRIFT_TILES: vec2<f32> = vec2<f32>(15.0, 6.0);
+
 // WarcraftXL's billow blend: fold each octave toward rounded puffs by `cotton` (0.6).
 fn billow(v: f32) -> f32 {
     return v + (1.0 - abs(2.0 * v - 1.0) - v) * 0.6;
@@ -256,7 +269,7 @@ fn cloud_billow(p: vec2<f32>, octaves: i32) -> f32 {
     var tot = 0.0;
     var q = p;
     for (var o = 0; o < octaves; o++) {
-        s += a * billow(value_noise_plane(q, u32(o) * 57u));
+        s += a * billow(value_noise_plane(q, u32(o) * 57u, CLOUD_TILE << u32(o)));
         tot += a;
         a *= 0.55;
         q = q * 2.0 + vec2<f32>(3.7, 1.9);
@@ -267,8 +280,14 @@ fn cloud_billow(p: vec2<f32>, octaves: i32) -> f32 {
 // The detail field in [0, 1] on the sheet: a broad warp twists the domain before the billowed fbm
 // is read, which turns thresholded blobs into cauliflower edges.
 fn cloud_detail(uv: vec2<f32>, t: f32, octaves: i32) -> f32 {
-    let p = uv * 36.0 + vec2<f32>(0.011, 0.0043) * t;
-    let w = vec2<f32>(value_noise_plane(p * 0.5, 173u), value_noise_plane(p * 0.5 + 5.3, 219u)) - 0.5;
+    // The drift modulo one tile: seamless at the wrap, and small enough to keep f32 precision.
+    let drift = fract(CLOUD_DRIFT_TILES * (t / SKY_WRAP)) * f32(CLOUD_TILE);
+    let p = uv * 36.0 + drift;
+    let half_tile = CLOUD_TILE / 2;
+    let w = vec2<f32>(
+        value_noise_plane(p * 0.5, 173u, half_tile),
+        value_noise_plane(p * 0.5 + 5.3, 219u, half_tile),
+    ) - 0.5;
     return cloud_billow(p + w * 2.6, octaves);
 }
 
diff --git a/crates/benilla-world/src/sky_fx.rs b/crates/benilla-world/src/sky_fx.rs
index 2c6accc4..4aead18a 100644
--- a/crates/benilla-world/src/sky_fx.rs
+++ b/crates/benilla-world/src/sky_fx.rs
@@ -47,8 +47,8 @@ impl SkyQuality {
 ///
 /// MONKEY (fix-sky): it wrapped hourly in f32, and every star's twinkle jumped at the wrap. Now it
 /// accumulates in f64 and wraps at [`SKY_CLOCK_WRAP_S`]; the shader's twinkle rates are whole
-/// cycles per wrap (seamless). The High cloud-detail drift still re-patterns at the wrap, once
-/// per 24 h of continuous play.
+/// cycles per wrap (seamless). MONKEY (polish): the High cloud detail tiles and drifts whole
+/// tiles per wrap (`CLOUD_TILE`, `CLOUD_DRIFT_TILES`), so it is seamless there too.
 #[derive(Resource, Default)]
 pub struct SkyClock {
     pub secs: f32,
@@ -175,6 +175,39 @@ fn tick_sky_clock(time: Res<Time>, mut clock: ResMut<SkyClock>) {
 mod tests {
     use super::*;
 
+    /// A `const NAME ... = <value>;` line of `sky_fx.wgsl`, as text.
+    fn wgsl_const(name: &str) -> String {
+        let src = include_str!("shaders/sky_fx.wgsl");
+        let line = src
+            .lines()
+            .find(|l| l.starts_with(&format!("const {name}:")))
+            .unwrap_or_else(|| panic!("no const {name}"));
+        let value = line.split('=').nth(1).unwrap().trim();
+        value.trim_end_matches(';').trim().to_string()
+    }
+
+    #[test]
+    fn the_shader_wrap_and_the_cloud_loop_match_the_clock() {
+        let wrap: f64 = wgsl_const("SKY_WRAP").parse().unwrap();
+        assert_eq!(wrap, SKY_CLOCK_WRAP_S);
+        // The loop needs an even tile (the half-rate warp tiles at half of it) and whole tiles of
+        // drift per wrap; the rates stay within 5% of WarcraftXL's 0.011 and 0.0043 cells/s.
+        let tile: i32 = wgsl_const("CLOUD_TILE").parse().unwrap();
+        assert!(tile > 0 && tile % 2 == 0);
+        let drift = wgsl_const("CLOUD_DRIFT_TILES");
+        let tiles: Vec<f64> = drift
+            .trim_start_matches("vec2<f32>(")
+            .trim_end_matches(')')
+            .split(',')
+            .map(|v| v.trim().parse().unwrap())
+            .collect();
+        for (n, rate) in tiles.iter().zip([0.011, 0.0043]) {
+            assert_eq!(n.fract(), 0.0, "whole tiles per wrap");
+            let actual = n * f64::from(tile) / SKY_CLOCK_WRAP_S;
+            assert!((actual / rate - 1.0).abs() < 0.05, "drift {actual} vs {rate}");
+        }
+    }
+
     #[test]
     fn glow_fades_out_below_the_horizon_and_under_cloud() {
         assert_eq!(glow_strength(-0.2, 0.0, 0.0), 0.0);