feat(particles): Étape 28 A — ParticlePool infrastructure (no driver)

Creates the GPU resource layer for particles per ARCHI §3.2/§6, without
simulation: the pool owns all buffers/pipeline/bind group but draws nothing
(indirect args zeroed → no-op) until a driver is attached (Étape B).

New:
- resources/particle.rs: Particle (80 B, #[repr(C)], no padding — D19),
  SIZE/ZERO consts + offset/layout unit tests
- shaders/particle_billboard.wgsl: camera-facing billboard, empty vertex
  layout (quad via vertex_index), instance slot via storage binding
  compact_index (D17), uv_rect atlas support (D15/D18)
- core/particles.rs: ParticlePoolConfig, BlendingMode, ParticleDriver trait
  (D3), ParticlePool (4 buffers + pipeline + bind group), default disc
  texture (D11), unit tests

Wired:
- Scene: SceneGpu keeps queue/sample_count; particle_pools registry +
  create_particle_pool() (default disc when no texture given)
- utils::conf PARTICLE_BILLBOARD_SHADER, module re-exports, prelude
- tests/wgsl_validate.rs: particle billboard naga validation (2 entry points)

Docs: DRAFT call-site/tree synced with the final code; AGENTS.md test count
(138) + wgpu 30 API drift gotcha (contents/DeviceExt/ALPHA_BLENDING,
DepthStencilState no Default, NonZero min_binding_size, const Zeroable).

cargo test -p wsg-lib: 138 pass (121 lib + 10 wgsl + 7), 0 warnings.

AGENTS.md

@@ -43,9 +43,10 @@ WGPU doesn't have a native "Context" object — this type groups them together f
 ## Gotchas
 - Rust 2024 edition is used. Ensure your Rust toolchain supports it (`rustup update`).
 - wgpu 30.0.0 is pinned in `lib/Cargo.toml`. The comment says "check the latest version" — verify compatibility before upgrading.
-- Cargo features gate primitives (`prim-*`, `all-prims` is default) and importers (`import-obj`, `import-gltf`); the `import` example is `required-features = ["import-obj"]`. 127 tests exist (`cargo test --workspace`).
+- Cargo features gate primitives (`prim-*`, `all-prims` is default) and importers (`import-obj`, `import-gltf`); the `import` example is `required-features = ["import-obj"]`. 138 tests exist (`cargo test --workspace`, incl. particle layout + billboard WGSL validation).
 - The workspace has no `[workspace.dependencies]` section. Dependencies are declared per-crate rather than centrally.
 - **WGSL `select` argument order** (cost us a day): `select(reject, accept, cond)` returns the **second** arg when `cond` is true — the reverse of HLSL's `select(trueVal, falseVal, cond)`. In `shaders/gpu_driven.wgsl` the cull pass must stay `select(0u, u32(flags.z), visible)` (visible ⇒ full count, culled ⇒ 0). Swapped args silently zero the counts of every visible entity → black window. See the GOTCHA comment at the top of that shader.
+- **wgpu 30 API drift** (verified this session): `BufferInitDescriptor` has a `contents: &[u8]` field (not `data`) and `create_buffer_init` comes from the `wgpu::util::DeviceExt` trait (import it, as in `mesh.rs`). `BlendState::ALPHA_BLENDING` is the alpha-blend constant (there is no `ALPHA`); `DepthStencilState` has **no** `Default` impl — write `stencil`/`bias` fields explicitly. `min_binding_size` is `Option<NonZero<u64>>`. bytemuck 1.25: `Zeroable::zeroed()` is not `const` (const traits unstable) — use a const literal for `ZERO`-style constants. For layout-offset tests prefer `std::mem::offset_of!` (stable 1.77, no unsafe).
 - **LOD UV blending: never fold integer-tile jumps, freeze seam twins instead** (cost us a day, 2026-09-23): a UV *seam* is two copies of the same 3-D point on integer-apart UVs (u=0/u=1 columns) — it is NOT a mesh edge, so the decimation must record the weld's refused pairs and **freeze** those twins (any edge touching one is excluded from the PQ). A co-facial edge spanning a whole tile (cone apex v=1 ↔ base v=0) is a *legit* chart span — the chart is bilinear, so the UVs **blend linearly** (fold the integer jump to zero and the apex UV smears down the cone side). And the attribute-aware weld refuses a Δ of *exactly* 0.5 (ambiguous: seam at its widest vs legit half-tile jump — the cone's u=1 column vs the cap-disc chart sits exactly there). See the comments in `geometry.rs` (`welded`, `Collapse::collapse_edge`) and the cone/seam regression tests.
 
 <!-- lean-ctx -->

docs/DRAFT.md

@@ -60,9 +60,9 @@ seront repris **après** le système de particules (phase 7).
 
 ```
 lib/
+└── src/
     ├── shaders/
     │   └── particle_billboard.wgsl   # NOUVEAU : vs_main + fs_main (pas de compute)
-└── src/
     ├── utils/
     │   └── conf.rs               # + pub const PARTICLE_BILLBOARD_SHADER (include_str!, pattern existant)
     ├── resources/
@@ -373,8 +373,7 @@ impl Scene {
         // Construire le pool (buffers + pipeline + bind group)
         let gpu = self.gpu();
         let pool = ParticlePool::new(
-            &gpu.device,
-            &gpu.queue,
+            gpu.device.as_ref(),
             gpu.format,
             gpu.sample_count,
             &config,

lib/src/core/mod.rs

@@ -19,6 +19,7 @@ pub mod geometry;
 pub mod hdr;
 pub mod lod;
 pub mod msaa;
+pub mod particles;
 pub mod renderer;
 pub mod shadow;
 pub mod transform;
@@ -34,6 +35,7 @@ pub use geometry::{BBox, Geometry, GeometryError};
 pub use hdr::ToneMapper;
 pub use lod::{lod_level, projected_radius_px};
 pub use msaa::MsaaConfig;
+pub use particles::{BlendingMode, ParticleDriver, ParticlePool, ParticlePoolConfig};
 pub use renderer::Renderer;
 pub use shadow::ShadowConfig;
 pub use transform::Transform;

lib/src/core/particles.rs

@@ -0,0 +1,361 @@
+//! # Particles — Pool Infrastructure (Étape 28 A)
+//!
+//! Implements the **pool** side of the Pool ≠ Driver architecture
+//! (ARCHI_PARTICULES §1–§3, §6–§8): the pool owns the GPU resources — the
+//! per-particle state buffer (80 B/slot, D15/D19), the compaction index buffer
+//! and the indirect draw args (D17), the per-pool camera-params uniform, the
+//! billboard render pipeline and its bind group — while *drivers* (GPU/CPU/Manual,
+//! Étapes B/C/D) provide the simulation logic through the [`ParticleDriver`] trait.
+//!
+//! A pool without an attached driver costs nothing at render time: its indirect
+//! args stay zero, so the draw is a no-op (D12). Pools are created through
+//! [`Scene::create_particle_pool`](crate::scene::Scene::create_particle_pool).
+
+use crate::pipeline::DEPTH_FORMAT;
+use crate::resources::Particle;
+use crate::utils::PARTICLE_BILLBOARD_SHADER;
+use wgpu::util::DeviceExt;
+
+/// Default pool capacity (particle slots).
+pub const DEFAULT_POOL_CAPACITY: u32 = 1024;
+
+/// Size in bytes of the per-pool camera-params uniform (view + proj — a prefix
+/// of `FrameUniforms`, so it can later be fed from the same buffer).
+pub const CAMERA_PARAMS_SIZE: u64 = 128;
+
+/// Blend mode frozen at pipeline creation (D9): one mode per pool.
+#[derive(Clone, Copy, PartialEq, Eq, Debug, Default)]
+pub enum BlendingMode {
+    /// Standard alpha blending (`SrcAlpha`/`OneMinusSrcAlpha` on color,
+    /// `One`/`OneMinusSrcAlpha` on alpha).
+    #[default]
+    Alpha,
+    /// Additive blending (`One`/`One` on both) — flames, sparks, glows.
+    Additive,
+}
+
+impl BlendingMode {
+    /// The wgpu blend state corresponding to this mode.
+    fn state(&self) -> wgpu::BlendState {
+        match self {
+            BlendingMode::Alpha => wgpu::BlendState::ALPHA_BLENDING,
+            BlendingMode::Additive => wgpu::BlendState::ADDITIVE,
+        }
+    }
+}
+
+/// Configuration for [`ParticlePool::new`].
+#[derive(Clone, Debug)]
+pub struct ParticlePoolConfig {
+    /// Pool capacity (particle slots). Default: [`DEFAULT_POOL_CAPACITY`].
+    pub max_count: u32,
+    /// Id of a texture already registered in the Scene. `None` → the built-in
+    /// 16×16 disc (D11).
+    pub texture: Option<String>,
+    /// Blend mode frozen at pipeline creation (D9).
+    pub blending: BlendingMode,
+}
+
+impl Default for ParticlePoolConfig {
+    fn default() -> Self {
+        Self {
+            max_count: DEFAULT_POOL_CAPACITY,
+            texture: None,
+            blending: BlendingMode::default(),
+        }
+    }
+}
+
+/// Common trait implemented by every driver type (ARCHI §4.1). The pool calls
+/// these methods in order each frame (wired into the render loop at Étape E).
+pub trait ParticleDriver: Send {
+    /// Called BEFORE the compute (if the driver is GPU) or before the draw.
+    /// The driver may write into the pool's buffers (spawns, CPU updates).
+    fn pre_compute(&mut self, queue: &wgpu::Queue, pool: &mut ParticlePool, dt: f32);
+    /// Called AFTER the compute (GPU driver only). Lets the driver update
+    /// post-simulation uniforms.
+    fn post_compute(&mut self, queue: &wgpu::Queue, pool: &mut ParticlePool);
+    /// Does the driver want a compute dispatch this frame?
+    fn needs_compute(&self) -> bool;
+    /// Does the driver want the pool drawn?
+    fn needs_draw(&self) -> bool;
+}
+
+/// The particle pool: all GPU resources needed to simulate and draw particles,
+/// without the simulation logic (which is the driver's job, D1).
+///
+/// `allow(dead_code)`: in Étape 28 A the resources are created but not yet consumed —
+/// Étapes B/E wire the driver, the compaction dispatch and the indirect draw into the
+/// render loop. The allow keeps the "zero warnings" acceptance criterion in the meantime.
+#[allow(dead_code)]
+pub struct ParticlePool {
+    /// Per-particle state: N × 80 B (D15/D19). STORAGE | COPY_DST, zeroed (all dead).
+    pub(crate) particle_data: wgpu::Buffer,
+    /// Compaction index: N × u32, one per slot (D17/D19). STORAGE | COPY_DST, zeroed.
+    pub(crate) compact_index: wgpu::Buffer,
+    /// Indirect draw args: 16 B (4 × u32) (D17). Zeroed → the draw is a no-op (D12).
+    pub(crate) indirect_args: wgpu::Buffer,
+    /// Per-pool camera params: 128 B (view + proj). UNIFORM | COPY_DST, owned by
+    /// the pool; the renderer writes it each frame (Étape E).
+    pub(crate) camera_params: wgpu::Buffer,
+    /// Billboard render pipeline (empty vertex layout, D6).
+    pub(crate) pipeline: wgpu::RenderPipeline,
+    /// The pool's bind group layout (5 bindings, ARCHI §6).
+    pub(crate) layout: wgpu::BindGroupLayout,
+    /// The single render bind group, built once at creation (D17/D19: the pool
+    /// owns all its buffers — self-contained).
+    pub(crate) bind_group: wgpu::BindGroup,
+    /// Texture view bound in the group (a scene texture, or the owned default disc, D11).
+    pub(crate) texture_view: wgpu::TextureView,
+    /// Pool capacity (particle slots).
+    pub max_count: u32,
+    /// Blend mode frozen at pipeline creation (D9).
+    pub blending: BlendingMode,
+    /// Attached driver (Étapes B/C/D). `None` → the pool is inactive (D12).
+    pub(crate) driver: Option<Box<dyn ParticleDriver>>,
+}
+
+impl ParticlePool {
+    /// Creates the pool's GPU resources: the 4 buffers, the shader module, the
+    /// bind group layout, the render pipeline (eagerly compiled) and the bind group.
+    pub fn new(
+        device: &wgpu::Device,
+        format: wgpu::TextureFormat,
+        sample_count: u32,
+        config: &ParticlePoolConfig,
+        texture_view: wgpu::TextureView,
+        sampler: wgpu::Sampler,
+    ) -> Self {
+        let n = config.max_count;
+
+        // --- Buffers (ARCHI §3.2/§6) — zeroed: all particles dead, empty draw args.
+        let particle_data = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+            label: Some("particle pool: particle_data"),
+            usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
+            contents: &vec![0u8; n as usize * Particle::SIZE as usize],
+        });
+        let compact_index = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+            label: Some("particle pool: compact_index"),
+            usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
+            contents: &vec![0u8; n as usize * 4],
+        });
+        let indirect_args = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+            label: Some("particle pool: indirect_args"),
+            usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
+            contents: &[0u8; 16],
+        });
+        let camera_params = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+            label: Some("particle pool: camera_params"),
+            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
+            contents: &vec![0u8; CAMERA_PARAMS_SIZE as usize],
+        });
+
+        // --- Shader + bind group layout (5 bindings, ARCHI §6).
+        let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
+            label: Some("particle_billboard"),
+            source: wgpu::ShaderSource::Wgsl(PARTICLE_BILLBOARD_SHADER.into()),
+        });
+        let layout = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
+            label: Some("particle pool: bind group layout"),
+            entries: &[
+                // 0: camera params (uniform, 128 B)
+                wgpu::BindGroupLayoutEntry {
+                    binding: 0,
+                    visibility: wgpu::ShaderStages::VERTEX,
+                    ty: wgpu::BindingType::Buffer {
+                        ty: wgpu::BufferBindingType::Uniform,
+                        has_dynamic_offset: false,
+                        min_binding_size: None,
+                    },
+                    count: None,
+                },
+                // 1: particle state (storage read-only)
+                wgpu::BindGroupLayoutEntry {
+                    binding: 1,
+                    visibility: wgpu::ShaderStages::VERTEX,
+                    ty: wgpu::BindingType::Buffer {
+                        ty: wgpu::BufferBindingType::Storage { read_only: true },
+                        has_dynamic_offset: false,
+                        min_binding_size: None,
+                    },
+                    count: None,
+                },
+                // 2: compaction index (storage read-only, D17/D19)
+                wgpu::BindGroupLayoutEntry {
+                    binding: 2,
+                    visibility: wgpu::ShaderStages::VERTEX,
+                    ty: wgpu::BindingType::Buffer {
+                        ty: wgpu::BufferBindingType::Storage { read_only: true },
+                        has_dynamic_offset: false,
+                        min_binding_size: None,
+                    },
+                    count: None,
+                },
+                // 3: sampler
+                wgpu::BindGroupLayoutEntry {
+                    binding: 3,
+                    visibility: wgpu::ShaderStages::FRAGMENT,
+                    ty: wgpu::BindingType::Sampler(wgpu::SamplerBindingType::Filtering),
+                    count: None,
+                },
+                // 4: particle texture
+                wgpu::BindGroupLayoutEntry {
+                    binding: 4,
+                    visibility: wgpu::ShaderStages::FRAGMENT,
+                    ty: wgpu::BindingType::Texture {
+                        sample_type: wgpu::TextureSampleType::Float { filterable: true },
+                        view_dimension: wgpu::TextureViewDimension::D2,
+                        multisampled: false,
+                    },
+                    count: None,
+                },
+            ],
+        });
+
+        // --- Render pipeline: EMPTY vertex layout (D6/D17/D19) — the quad is
+        //     generated in the shader (QUAD[vi]), the instance slot comes from
+        //     storage binding 2.
+        let pipeline_layout = device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
+            label: Some("particle pool: pipeline layout"),
+            bind_group_layouts: &[Some(&layout)],
+            immediate_size: 0,
+        });
+        let blend = config.blending.state();
+        let pipeline = device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
+            label: Some("particle pool: billboard pipeline"),
+            layout: Some(&pipeline_layout),
+            vertex: wgpu::VertexState {
+                module: &shader,
+                entry_point: Some("vs_main"),
+                compilation_options: Default::default(),
+                buffers: &[], // empty layout (D17/D19)
+            },
+            fragment: Some(wgpu::FragmentState {
+                module: &shader,
+                entry_point: Some("fs_main"),
+                compilation_options: Default::default(),
+                targets: &[Some(wgpu::ColorTargetState {
+                    format,
+                    blend: Some(blend),
+                    write_mask: wgpu::ColorWrites::ALL,
+                })],
+            }),
+            primitive: wgpu::PrimitiveState {
+                topology: wgpu::PrimitiveTopology::TriangleList,
+                ..Default::default()
+            },
+            depth_stencil: Some(wgpu::DepthStencilState {
+                format: DEPTH_FORMAT,
+                depth_write_enabled: Some(false), // D10
+                depth_compare: Some(wgpu::CompareFunction::LessEqual),
+                stencil: wgpu::StencilState::default(),
+                bias: wgpu::DepthBiasState::default(),
+            }),
+            multisample: wgpu::MultisampleState {
+                count: sample_count,
+                ..Default::default()
+            },
+            multiview_mask: None,
+            cache: None,
+        });
+
+        // --- Bind group (built once: the pool owns all its buffers, D17/D19).
+        let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
+            label: Some("particle pool: bind group"),
+            layout: &layout,
+            entries: &[
+                wgpu::BindGroupEntry {
+                    binding: 0,
+                    resource: camera_params.as_entire_binding(),
+                },
+                wgpu::BindGroupEntry {
+                    binding: 1,
+                    resource: particle_data.as_entire_binding(),
+                },
+                wgpu::BindGroupEntry {
+                    binding: 2,
+                    resource: compact_index.as_entire_binding(),
+                },
+                wgpu::BindGroupEntry {
+                    binding: 3,
+                    resource: wgpu::BindingResource::Sampler(&sampler),
+                },
+                wgpu::BindGroupEntry {
+                    binding: 4,
+                    resource: wgpu::BindingResource::TextureView(&texture_view),
+                },
+            ],
+        });
+
+        Self {
+            particle_data,
+            compact_index,
+            indirect_args,
+            camera_params,
+            pipeline,
+            layout,
+            bind_group,
+            texture_view,
+            max_count: n,
+            blending: config.blending,
+            driver: None,
+        }
+    }
+}
+
+/// RGBA bytes of the built-in 16×16 soft disc (D11), used when a pool has no texture.
+/// A radial gradient (opaque center → transparent edge), white.
+pub(crate) fn default_disc_rgba() -> Vec<u8> {
+    let size = 16;
+    let mut data = vec![0u8; size * size * 4];
+    let center = (size as f32 - 1.0) / 2.0;
+    for y in 0..size {
+        for x in 0..size {
+            let dx = (x as f32 - center) / center;
+            let dy = (y as f32 - center) / center;
+            let dist = (dx * dx + dy * dy).sqrt();
+            let alpha = ((1.0 - dist).clamp(0.0, 1.0) * 255.0) as u8;
+            let i = (y * size + x) * 4;
+            data[i] = 255;
+            data[i + 1] = 255;
+            data[i + 2] = 255;
+            data[i + 3] = alpha;
+        }
+    }
+    data
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn pool_config_default() {
+        let c = ParticlePoolConfig::default();
+        assert_eq!(c.max_count, DEFAULT_POOL_CAPACITY);
+        assert!(c.texture.is_none());
+        assert_eq!(c.blending, BlendingMode::Alpha);
+    }
+
+    #[test]
+    fn default_disc_size() {
+        assert_eq!(default_disc_rgba().len(), 16 * 16 * 4);
+    }
+
+    #[test]
+    fn default_disc_center_opaque() {
+        let d = default_disc_rgba();
+        // 16×16 is even: the exact center falls between the four middle pixels, so the
+        // highest alpha is 1 - sqrt(2)/15 ≈ 0.905 → 230, not 255. The soft disc must still
+        // be (nearly) opaque at its core.
+        let i = (7 * 16 + 7) * 4;
+        assert!(d[i + 3] >= 200, "center alpha was {}", d[i + 3]);
+    }
+
+    #[test]
+    fn default_disc_corner_transparent() {
+        let d = default_disc_rgba();
+        assert_eq!(d[3], 0); // pixel (0, 0)
+        assert_eq!(d[(15 * 16 + 15) * 4 + 3], 0); // pixel (15, 15)
+    }
+}

lib/src/prelude.rs

@@ -15,8 +15,11 @@
 // Core types
 pub use crate::core::geometry::{BBox, Geometry};
 pub use crate::core::transform::Transform;
-pub use crate::core::{BloomConfig, DoFConfig, FogConfig, FogMode, MsaaConfig, ShadowConfig, ToneMapper};
-pub use crate::resources::Material;
+pub use crate::core::{
+    BlendingMode, BloomConfig, DoFConfig, FogConfig, FogMode, MsaaConfig, ParticleDriver,
+    ParticlePool, ParticlePoolConfig, ShadowConfig, ToneMapper,
+};
+pub use crate::resources::{Material, Particle};
 
 // Camera
 pub use crate::camera::{Camera, CameraController};

lib/src/resources/mod.rs

@@ -9,6 +9,7 @@
 
 pub mod material;
 pub mod mesh;
+pub mod particle;
 pub mod texture;
 pub mod uniform;
 pub mod vertex;
@@ -16,6 +17,7 @@ pub mod vertex;
 // Re-exports
 pub use material::Material;
 pub use mesh::{LodMode, Mesh, PackError};
+pub use particle::Particle;
 pub use texture::{Texture, TextureError};
 pub use uniform::{
     BBOX_SLOT_SIZE, BBoxSlot, CULL_UNIFORMS_SIZE, CullUniforms, DRAW_SLOT_SIZE, DrawSlot,

lib/src/resources/particle.rs

@@ -0,0 +1,104 @@
+//! # Particle Resource — Per-Particle GPU State (Étape 28)
+//!
+//! CPU-side mirror of the WGSL `Particle` struct (particle system, Step 28 A): one
+//! **80-byte** flat record per slot in the pool's storage buffer (ARCHI_PARTICULES §2,
+//! decisions D15/D19).
+//!
+//! The layout is **padding-free**: the buffer lives in WGSL *storage* space where
+//! `vec3<f32>`/`vec4<f32>` have alignment 4 (uniform space would enforce alignment 16),
+//! so the Rust `#[repr(C)]` layout matches the WGSL struct field-for-field with no
+//! padding. The size and the per-field offsets are pinned by tests below.
+
+use bytemuck::{Pod, Zeroable};
+
+/// Per-particle state, 80 bytes — the CPU mirror of the WGSL `Particle` struct
+/// (`shaders/particle_billboard.wgsl`, and later `particle_update.wgsl`).
+///
+/// A slot is *dead* when `life == 0.0` (a fresh pool buffer is all-dead).
+/// Drivers (GPU/CPU/Manual) write these records; the billboard render pass reads them.
+#[repr(C)]
+#[derive(Copy, Clone, Debug, PartialEq, Pod, Zeroable, Default)]
+pub struct Particle {
+    /// World position (xyz). Offset 0.
+    pub pos: [f32; 3],
+    /// Velocity (xyz). Offset 12.
+    pub vel: [f32; 3],
+    /// Remaining lifetime (seconds). Offset 24. `0.0` = dead.
+    pub life: f32,
+    /// Initial lifetime (seconds) — normalizes the fade (`life / max_life`). Offset 28.
+    pub max_life: f32,
+    /// Current size (world units). Offset 32.
+    pub size: f32,
+    /// Size growth (units per second; positive grows, negative shrinks). Offset 36.
+    pub size_growth: f32,
+    /// Current 2D rotation (radians). Offset 40.
+    pub angle: f32,
+    /// Angular velocity (radians per second). Offset 44.
+    pub angular_vel: f32,
+    /// RGBA color (the alpha component is driven by the integrator, D18). Offset 48.
+    pub color: [f32; 4],
+    /// UV rect `(ox, oy, sx, sy)` — atlas support (D15). Offset 64.
+    pub uv_rect: [f32; 4],
+}
+
+impl Particle {
+    /// Size of one slot in the pool's storage buffer: **80 bytes** (pinned by tests).
+    pub const SIZE: u64 = std::mem::size_of::<Self>() as u64;
+
+    /// The all-dead particle (`life == 0.0`). A fresh pool buffer is this record
+    /// repeated N times.
+    pub const ZERO: Self = Self {
+        pos: [0.0; 3],
+        vel: [0.0; 3],
+        life: 0.0,
+        max_life: 0.0,
+        size: 0.0,
+        size_growth: 0.0,
+        angle: 0.0,
+        angular_vel: 0.0,
+        color: [0.0; 4],
+        uv_rect: [0.0; 4],
+    };
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::mem::size_of;
+
+    /// D15/D19: the layout is exactly 80 bytes, without any padding.
+    #[test]
+    fn particle_size_is_80() {
+        assert_eq!(size_of::<Particle>(), 80);
+        assert_eq!(Particle::SIZE, 80);
+    }
+
+    /// Storage space: the maximum field alignment is 4 (vec3/vec4 are align 4 in
+    /// storage, not uniform space) — so no padding is needed anywhere.
+    #[test]
+    fn particle_align_is_4() {
+        assert_eq!(std::mem::align_of::<Particle>(), 4);
+    }
+
+    /// The offsets must match the WGSL `Particle` struct exactly (storage space,
+    /// no padding). A drift here silently corrupts the whole pool on the GPU.
+    #[test]
+    fn particle_field_offsets() {
+        assert_eq!(std::mem::offset_of!(Particle, pos), 0);
+        assert_eq!(std::mem::offset_of!(Particle, vel), 12);
+        assert_eq!(std::mem::offset_of!(Particle, life), 24);
+        assert_eq!(std::mem::offset_of!(Particle, max_life), 28);
+        assert_eq!(std::mem::offset_of!(Particle, size), 32);
+        assert_eq!(std::mem::offset_of!(Particle, size_growth), 36);
+        assert_eq!(std::mem::offset_of!(Particle, angle), 40);
+        assert_eq!(std::mem::offset_of!(Particle, angular_vel), 44);
+        assert_eq!(std::mem::offset_of!(Particle, color), 48);
+        assert_eq!(std::mem::offset_of!(Particle, uv_rect), 64);
+    }
+
+    #[test]
+    fn particle_zero_is_dead() {
+        assert_eq!(Particle::ZERO.life, 0.0);
+        assert_eq!(Particle::ZERO, Particle::default());
+    }
+}

lib/src/scene/scene.rs

@@ -15,7 +15,9 @@
 //! instead of `App`. It can therefore build materials and meshes itself (`add_material_shader`, `create_mesh`) and inject
 //! a default material for meshes that carry none (`default_material`).
 
-use crate::core::{Geometry, Transform};
+use crate::core::{
+    particles::{default_disc_rgba, ParticlePool, ParticlePoolConfig}, Geometry, Transform,
+};
 use crate::pipeline::PipelineCache;
 use crate::camera::Camera; use crate::lights::Lights; use crate::resources::{BBoxSlot, Material, Mesh, Texture, TransformSlot};
 use crate::scene::Entity;
@@ -31,8 +33,13 @@ use std::sync::Arc;
 struct SceneGpu {
     /// Shared GPU device used to create mesh buffers and compile pipelines.
     device: Arc<wgpu::Device>,
+    /// GPU command queue (uploads). Kept for resources built at creation time
+    /// (e.g. the default particle disc texture, Étape 28 A).
+    queue: wgpu::Queue,
     /// Surface texture output format, required to build fragment pipelines.
     format: wgpu::TextureFormat,
+    /// MSAA sample count (pipeline multisample state).
+    sample_count: u32,
     /// Shader compilation cache: compiles/caches RenderPipelines keyed by shader_id + format.
     cache: RefCell<PipelineCache>,
 }
@@ -82,6 +89,9 @@ pub struct Scene {
     meshes: HashMap<String, Arc<Mesh>>,
     /// Map of material identifiers to owned `Arc<Material>` instances. Populated via `add_material()`.
     materials: HashMap<String, Arc<Material>>,
+    /// Map of particle pool identifiers to owned `Arc<ParticlePool>` instances (Étape 28 A).
+    /// Populated via `create_particle_pool()`.
+    particle_pools: HashMap<String, Arc<ParticlePool>>,
     /// Map of diffuse texture identifiers to owned `Arc<Texture>` instances (Step 10, D4).
     /// Populated via `add_texture()`; materials reference them via `add_material_texture()` by id.
     textures: HashMap<String, Arc<Texture>>,
@@ -124,6 +134,7 @@ impl Scene {
             meshes: HashMap::new(),
             materials: HashMap::new(),
             textures: HashMap::new(),
+            particle_pools: HashMap::new(),
             entities: HashMap::new(),
             camera: Camera::default(),
             gpu: None,
@@ -153,7 +164,9 @@ impl Scene {
         let cache = PipelineCache::new(device.clone(), queue.clone(), sample_count);
         self.gpu = Some(SceneGpu {
             device,
+            queue,
             format,
+            sample_count,
             cache: RefCell::new(cache),
         });
         self
@@ -262,6 +275,57 @@ impl Scene {
         self.textures.get(id)
     }
 
+    /// Étape 28 A : crée un pool de particules (infra GPU, sans driver).
+    ///
+    /// Le pool possède ses buffers (état 80 B/slot, index de compaction, args indirect,
+    /// camera params), son pipeline billboard et son bind group (ARCHI §3.2/§6).
+    /// `config.texture` est l'id d'une texture enregistrée via `add_texture` ;
+    /// `None` → disque 16×16 intégré (D11). Un pool sans driver ne draw rien (D12).
+    pub fn create_particle_pool(
+        &mut self,
+        id: &str,
+        config: ParticlePoolConfig,
+    ) -> Result<(), String> {
+        if self.particle_pools.contains_key(id) {
+            return Err(format!("particle pool '{id}' already exists"));
+        }
+        let gpu = self.gpu();
+        let (texture_view, sampler) = match &config.texture {
+            Some(tex_id) => {
+                let tex = self
+                    .textures
+                    .get(tex_id)
+                    .ok_or_else(|| {
+                        format!("texture '{tex_id}' not found — register it via add_texture first")
+                    })?;
+                (tex.view.clone(), tex.sampler.clone())
+            }
+            None => {
+                let disc = Texture::from_rgba8(
+                    gpu.device.as_ref(),
+                    &gpu.queue,
+                    16,
+                    16,
+                    &default_disc_rgba(),
+                    "particle default disc",
+                )
+                .map_err(|e| format!("failed to build default disc texture: {e}"))?;
+                (disc.view, disc.sampler)
+            }
+        };
+        let pool = ParticlePool::new(
+            gpu.device.as_ref(),
+            gpu.format,
+            gpu.sample_count,
+            &config,
+            texture_view,
+            sampler,
+        );
+        self.particle_pools
+            .insert(id.to_string(), Arc::new(pool));
+        Ok(())
+    }
+
     /// Builds and registers a Material from a shader id **and** a diffuse texture registered via
     /// [`Scene::add_texture`]. The material samples `texture_id` (Step 10, D4). Returns Ok(id) or
     /// Err(String) if the material id exists or the texture id does not. Inputs: id (material id to

lib/src/shaders/particle_billboard.wgsl

@@ -0,0 +1,86 @@
+// Particle billboard shader (Étape 28 A) — vertex + fragment, no compute yet.
+//
+// EMPTY vertex layout: the quad is generated in the shader (QUAD[vi], 6 vertices =
+// 2 triangles) and each instance's pool slot arrives via a STORAGE binding
+// (`compact_index[ii]`), not as a vertex attribute (ARCHI D17/D19 — WebGPU-safe,
+// same pattern as the TM fullscreen triangle).
+// The instance count comes from the indirect args written by the compaction (D17):
+// no early-out, no count uniform — a zeroed args buffer makes the draw a no-op (D12).
+
+struct Particle {  // 80 B — storage space: vec3/vec4 align 4, no padding (D19)
+    pos: vec3<f32>,          // offset 0
+    vel: vec3<f32>,          // offset 12
+    life: f32,               // offset 24 — 0.0 = dead
+    max_life: f32,           // offset 28
+    size: f32,               // offset 32
+    size_growth: f32,        // offset 36
+    angle: f32,              // offset 40
+    angular_vel: f32,        // offset 44
+    color: vec4<f32>,        // offset 48
+    uv_rect: vec4<f32>,      // offset 64 — (D15) (ox, oy, sx, sy)
+}
+
+struct CameraParams {  // 128 B — prefix of FrameUniforms (view + proj)
+    view: mat4x4<f32>,
+    proj: mat4x4<f32>,
+}
+
+struct VsOut {
+    @builtin(position) clip: vec4<f32>,
+    @location(0) frag_color: vec4<f32>,
+    @location(1) uv: vec2<f32>,
+}
+
+@group(0) @binding(0) var<uniform> camera: CameraParams;
+@group(0) @binding(1) var<storage, read> particles: array<Particle>;
+@group(0) @binding(2) var<storage, read> compact_index: array<u32>;  // (D17/D19)
+
+// 6 entries = 2 triangles (0-1-2, 3-4-5) forming one quad.
+// Without an index buffer, draw(4, n) would yield a single triangle + 1 orphan vertex.
+const QUAD: array<vec2<f32>, 6> = array<vec2<f32>, 6>(
+    vec2(-0.5, -0.5),
+    vec2( 0.5, -0.5),
+    vec2( 0.5,  0.5),
+    vec2(-0.5, -0.5),
+    vec2( 0.5,  0.5),
+    vec2(-0.5,  0.5),
+);
+
+@vertex
+fn vs_main(
+    @builtin(vertex_index) vi: u32,
+    @builtin(instance_index) ii: u32,
+) -> VsOut {
+    var out: VsOut;
+
+    // Instance slot via storage (D17/D19) — no early-out: the instance count is
+    // exactly the number of alive particles (indirect args).
+    let slot = compact_index[ii];
+    let p = particles[slot];
+
+    let q = QUAD[vi];
+
+    // 2D rotation in the billboard plane
+    let c = cos(p.angle);
+    let s = sin(p.angle);
+    let rot = vec2(q.x * c - q.y * s, q.x * s + q.y * c) * p.size;
+
+    // Camera-facing axes (columns 0 and 1 of the view matrix)
+    let right = vec3(camera.view[0][0], camera.view[1][0], camera.view[2][0]);
+    let up    = vec3(camera.view[0][1], camera.view[1][1], camera.view[2][1]);
+
+    let world = p.pos + right * rot.x + up * rot.y;
+    out.clip = camera.proj * camera.view * vec4(world, 1.0);
+    out.frag_color = p.color;
+    out.uv = p.uv_rect.xy + (q + vec2(0.5)) * p.uv_rect.zw;  // (D15)
+    return out;
+}
+
+@group(0) @binding(3) var samp: sampler;
+@group(0) @binding(4) var tex: texture_2d<f32>;
+
+@fragment
+fn fs_main(in: VsOut) -> @location(0) vec4<f32> {
+    let t = textureSample(tex, samp, in.uv);
+    return in.frag_color * t;
+}

lib/src/utils/conf.rs

@@ -64,6 +64,12 @@ pub const DOF_COC_SHADER: &str = include_str!("../shaders/dof_coc.wgsl");
 /// DoF blur shader (Étape 26).
 pub const DOF_BLUR_SHADER: &str = include_str!("../shaders/dof_blur.wgsl");
 
+/// Particle billboard shader (Étape 28 A), embedded at compile time.
+/// Carries one vertex entry point (`vs_main` — EMPTY layout: quad via
+/// `@builtin(vertex_index)`, instance slot via storage binding, D17/D19) and one fragment
+/// entry point (`fs_main`). Compiled directly by `ParticlePool` (library-internal pipeline).
+pub const PARTICLE_BILLBOARD_SHADER: &str = include_str!("../shaders/particle_billboard.wgsl");
+
 /// Fixed capacity of the GPU-driven entity slot buffers (Phase 3). The transform, matrix, bbox and
 /// indirect-draw-args buffers are all sized to this capacity and allocated once; per frame the CPU
 /// rewrites only the transform slots and the cull uniforms.

lib/src/utils/mod.rs

@@ -14,7 +14,7 @@ pub mod error;
 
 // Re-exports
 pub use conf::{
-    SHADOW_MAP_SIZE, SHADOW_SCENE_CENTER, SHADOW_SCENE_RADIUS, SHADOW_SHADER, SHADOW_SHADER_PATH,
-    STANDARD_SHADER, STANDARD_SHADER_PATH,
+    PARTICLE_BILLBOARD_SHADER, SHADOW_MAP_SIZE, SHADOW_SCENE_CENTER, SHADOW_SCENE_RADIUS,
+    SHADOW_SHADER, SHADOW_SHADER_PATH, STANDARD_SHADER, STANDARD_SHADER_PATH,
 };
 pub use error::WsgError;

lib/tests/wgsl_validate.rs

@@ -86,6 +86,37 @@ fn gpu_driven_shader_is_valid_wgsl() {
     );
 }
 
+/// Parses and fully validates the embedded `particle_billboard.wgsl` shader (Étape 28 A) via naga.
+/// The particle pool compiles it into one `RenderPipeline` (empty vertex layout — quad via
+/// `@builtin(vertex_index)`, instance slot via storage binding, D17/D19), so this offline
+/// validation is the guarantee of its validity. The contract expects two entry points:
+/// `vs_main` + `fs_main` (no compute in this step).
+#[test]
+fn particle_billboard_shader_is_valid_wgsl() {
+    let src = include_str!("../src/shaders/particle_billboard.wgsl");
+    let module = naga::front::wgsl::parse_str(src)
+        .unwrap_or_else(|e| panic!("particle_billboard.wgsl: parsing error: {e:?}"));
+
+    let mut validator = naga::valid::Validator::new(
+        naga::valid::ValidationFlags::all(),
+        naga::valid::Capabilities::all(),
+    );
+    validator
+        .validate(&module)
+        .unwrap_or_else(|e| panic!("particle_billboard.wgsl: validation failed: {e:?}"));
+
+    let entry_names: Vec<&str> = module
+        .entry_points
+        .iter()
+        .map(|ep| ep.name.as_str())
+        .collect();
+    assert_eq!(
+        entry_names,
+        vec!["vs_main", "fs_main"],
+        "the two entry points (vs + fs) are expected"
+    );
+}
+
 /// Parses and fully validates the embedded `tonemap.wgsl` shader (Étape 20) via naga.
 /// The renderer compiles it into one `RenderPipeline` (vertex `vs_main` + one of the two
 /// fragment entry points `fs_aces` / `fs_reinhard`), so this offline validation is the