rustial_renderer_wgpu/

renderer.rs

// ---------------------------------------------------------------------------
//! # Top-level WGPU renderer
//!
//! [`WgpuMapRenderer`] is the single entry-point for drawing the map with
//! a pure-WGPU backend.  The host application creates it once during
//! initialisation and calls [`render`](WgpuMapRenderer::render) (or the
//! richer [`render_full`](WgpuMapRenderer::render_full)) each frame.
//!
//! ## Ownership model
//!
//! The renderer **does not** own the `wgpu::Device`, `wgpu::Queue`, or
//! surface.  These are provided by the host (typically via winit + WGPU
//! surface setup) and passed in by reference.  This keeps the renderer
//! framework-agnostic: it works identically inside a winit event loop,
//! an egui integration, or a headless offscreen test.
//!
//! ## Camera-relative rendering
//!
//! All world-space positions (tiles, terrain, vectors, models) are
#![allow(clippy::many_single_char_names)]
//! transformed to *camera-relative* f32 coordinates before upload.
//! The camera's world-space origin is subtracted on the CPU so that
//! vertices near the camera are close to `(0, 0, 0)`.  This avoids
//! catastrophic f32 precision loss when the camera is far from the
//! Web Mercator origin (e.g. at longitude 170 degrees, x ~ 19 million meters).
//!
//! The view-projection matrix is similarly computed relative to the
//! camera origin and uploaded as a single shared uniform buffer.
//!
//! ## GPU resource layout
//!
//! | Resource | Lifetime | Notes |
//! |----------|----------|-------|
//! | Pipelines (tile, terrain, vector, model) | Renderer | Created once |
//! | Uniform buffer + bind groups | Renderer | One buffer, four BGs (shared layout) |
//! | Tile atlas (`TileAtlas`) | Renderer | Grows lazily, evicted per-frame |
//! | Page bind groups | Renderer | Rebuilt when atlas pages are added |
//! | Tile batch buffers | Cached | Invalidated when visible set or camera origin changes |
//! | Vector batch buffers | Cached | Invalidated when mesh data or camera origin changes |
//! | Terrain batch buffers (fallback) | Per-frame | Only used for non-standard projections |
//! | Model mesh buffers | Cached | Keyed by `ModelMeshKey` fingerprint |
//! | Model transform buffers | Cached | Invalidated when instance set or camera origin changes |
//! | Shared terrain grid meshes | Cached | One per resolution, reused across all tiles |
//! | Elevation textures | Cached | Per-tile R32Float, invalidated on generation change |
//! | Depth texture | Renderer | Recreated on `resize()` |
//! | Sampler | Renderer | Bilinear, clamp-to-edge |
//!
//! ## Frame lifecycle
//!
//! ```text
//! render_full(params)
//!   1. Upload view-proj uniform
//!   2. Upload new tile textures into the atlas
//!   3. Mark visible tiles + terrain tiles as used
//!   4. Cache model mesh GPU buffers (avoids re-upload)
//!   5. Build batched geometry (tiles, terrain, vectors)
//!   6. Begin render pass (clear colour + depth)
//!      a. Draw terrain batches  -- OR --  flat tile batches
//!      b. Draw vector overlays
//!      c. Draw 3D model instances
//!   7. Submit command buffer
//!   8. Atlas end-of-frame eviction
//! ```
// ---------------------------------------------------------------------------

use crate::gpu::batch::{
    build_circle_batch, build_fill_batch, build_fill_extrusion_batch, build_fill_pattern_batch,
    build_heatmap_batch, build_hillshade_batches, build_line_batch, build_line_pattern_batch,
    build_placeholder_batches, build_symbol_batch, build_terrain_batches, build_tile_batches,
    build_vector_batch, find_terrain_texture_actual, CircleBatchEntry, FillBatchEntry,
    FillExtrusionBatchEntry, FillPatternBatchEntry, HeatmapBatchEntry, HillshadeBatch,
    LineBatchEntry, LinePatternBatchEntry, SymbolBatchEntry, TerrainBatch, TilePageBatches,
    VectorBatchEntry,
};
use crate::gpu::column_vertex::{ColumnInstanceData, ColumnVertex};
use crate::gpu::depth::create_depth_texture;
use crate::gpu::grid_extrusion_vertex::GridExtrusionVertex;
use crate::gpu::grid_scalar_vertex::GridScalarVertex;
use crate::gpu::image_overlay_vertex::ImageOverlayVertex;
use crate::gpu::model_vertex::ModelVertex;
use crate::gpu::terrain_buffers::TerrainInteractionBuffers;
use crate::gpu::terrain_grid_vertex::TerrainGridVertex;
use crate::gpu::tile_atlas::TileAtlas;
use crate::painter::{PainterPass, PainterPlan};
use crate::pipeline::circle_pipeline::CirclePipeline;
use crate::pipeline::column_pipeline::ColumnPipeline;
use crate::pipeline::fill_extrusion_pipeline::FillExtrusionPipeline;
use crate::pipeline::fill_pattern_pipeline::FillPatternPipeline;
use crate::pipeline::fill_pipeline::FillPipeline;
use crate::pipeline::grid_extrusion_pipeline::GridExtrusionPipeline;
use crate::pipeline::grid_scalar_pipeline::GridScalarPipeline;
use crate::pipeline::heatmap_colormap_pipeline::HeatmapColormapPipeline;
use crate::pipeline::heatmap_pipeline::HeatmapPipeline;
use crate::pipeline::hillshade_pipeline::HillshadePipeline;
use crate::pipeline::image_overlay_pipeline::ImageOverlayPipeline;
use crate::pipeline::line_pattern_pipeline::LinePatternPipeline;
use crate::pipeline::line_pipeline::LinePipeline;
use crate::pipeline::model_pipeline::ModelPipeline;
use crate::pipeline::sky_pipeline::{SkyPipeline, SkyUniform};
use crate::pipeline::symbol_pipeline::SymbolPipeline;
use crate::pipeline::terrain_data_pipeline::TerrainDataPipeline;
use crate::pipeline::terrain_pipeline::TerrainPipeline;
use crate::pipeline::tile_pipeline::TilePipeline;
use crate::pipeline::uniforms::ViewProjUniform;
use crate::pipeline::vector_pipeline::VectorPipeline;
use glam::{DVec3, Mat4};
use rustial_engine as rustial_math;
use rustial_engine::TileId;
use rustial_engine::{
    materialize_terrain_mesh, DecodedImage, LayerId, MapState, ModelInstance, TerrainMeshData,
    TileData, VectorMeshData, VectorRenderMode, VisibleTile, VisualizationOverlay,
};
use std::sync::Arc;
use wgpu::util::DeviceExt;

#[repr(C)]
#[derive(Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
struct TerrainTileUniform {
    geo_bounds: [f32; 4],
    scene_origin: [f32; 4],
    elev_params: [f32; 4],
    elev_region: [f32; 4],
}

struct SharedTerrainGridMesh {
    vertex_buffer: wgpu::Buffer,
    index_buffer: wgpu::Buffer,
    index_count: u32,
}

struct CachedHeightTexture {
    generation: u64,
    view: wgpu::TextureView,
}

#[repr(C)]
#[derive(Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
struct GridScalarUniform {
    origin_counts: [f32; 4],
    grid_params: [f32; 4],
    scene_origin: [f32; 4],
    value_params: [f32; 4],
    base_altitude: [f32; 4],
}

struct SharedColumnMesh {
    vertex_buffer: wgpu::Buffer,
    index_buffer: wgpu::Buffer,
    index_count: u32,
}

struct CachedGridScalarOverlay {
    vertex_buffer: wgpu::Buffer,
    index_buffer: wgpu::Buffer,
    index_count: u32,
    vertex_count: usize,
    #[allow(dead_code)]
    uniform_buffer: wgpu::Buffer,
    bind_group: wgpu::BindGroup,
    #[allow(dead_code)]
    scalar_texture: wgpu::Texture,
    #[allow(dead_code)]
    ramp_texture: wgpu::Texture,
    generation: u64,
    value_generation: u64,
    ramp_fingerprint: u64,
    grid_fingerprint: u64,
    terrain_fingerprint: u64,
    projection: rustial_engine::CameraProjection,
    origin_key: [i64; 3],
}

struct CachedGridExtrusionOverlay {
    vertex_buffer: wgpu::Buffer,
    index_buffer: wgpu::Buffer,
    index_count: u32,
    vertex_count: usize,
    generation: u64,
    value_generation: u64,
    origin_key: [i64; 3],
    grid_fingerprint: u64,
    params_fingerprint: u64,
    ramp_fingerprint: u64,
    terrain_fingerprint: u64,
}

struct CachedColumnOverlay {
    instance_buffer: wgpu::Buffer,
    instance_count: u32,
    generation: u64,
    origin_key: [i64; 3],
    columns_fingerprint: u64,
    ramp_fingerprint: u64,
    instance_data: Vec<ColumnInstanceData>,
}

struct CachedPointCloudOverlay {
    instance_buffer: wgpu::Buffer,
    instance_count: u32,
    generation: u64,
    origin_key: [i64; 3],
    points_fingerprint: u64,
    ramp_fingerprint: u64,
    instance_data: Vec<ColumnInstanceData>,
}

/// Per-frame visualization cache activity recorded during the last render.
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
pub struct VisualizationPerfStats {
    /// Number of grid-scalar overlay cache rebuilds.
    pub grid_scalar_rebuilds: u32,
    /// Number of grid-scalar value-texture updates.
    pub grid_scalar_value_updates: u32,
    /// Number of grid-extrusion overlay rebuilds.
    pub grid_extrusion_rebuilds: u32,
    /// Number of grid-extrusion vertex-buffer updates.
    pub grid_extrusion_value_updates: u32,
    /// Number of column overlay rebuilds.
    pub column_rebuilds: u32,
    /// Number of partial column buffer writes.
    pub column_partial_writes: u32,
    /// Number of changed contiguous column ranges written this frame.
    pub column_partial_write_ranges: u32,
    /// Number of point-cloud overlay rebuilds.
    pub point_cloud_rebuilds: u32,
    /// Number of point-cloud partial retained writes.
    pub point_cloud_partial_writes: u32,
    /// Number of changed contiguous point-cloud ranges written this frame.
    pub point_cloud_partial_write_ranges: u32,
}

/// Cached per-tile terrain uniform buffer + bind group.
///
/// Matches MapLibre's approach of retaining per-tile GPU state across
/// frames and only recreating when the tile's data or the camera origin
/// changes.  This avoids allocating a fresh `wgpu::Buffer` and
/// `wgpu::BindGroup` for every terrain tile on every frame.
struct CachedTerrainTileBind {
    #[allow(dead_code)]
    uniform_buffer: wgpu::Buffer,
    bind_group: wgpu::BindGroup,
    /// Quantised scene origin used when this entry was created.
    origin_key: [i64; 3],
    /// Elevation data generation when this entry was created.
    generation: u64,
}

/// Cache key for per-tile terrain uniform/bind-group entries.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
struct TerrainTileBindKey {
    tile: TileId,
    /// Which pipeline family created this entry.
    pipeline: TerrainPipelineKind,
}

/// Distinguish terrain vs terrain-data vs hillshade pipeline bind groups.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
enum TerrainPipelineKind {
    Terrain,
    TerrainData,
    Hillshade,
}

/// Dirty-tracking state for the terrain-data interaction pass.
///
/// Mirrors MapLibre's `terrainFacilitator` pattern: the depth/coordinate
/// framebuffers are only redrawn when the VP matrix or the visible terrain
/// set has changed, rather than unconditionally every frame.
struct TerrainDataDirtyState {
    /// Whether an explicit dirty flag has been set (e.g. after resize).
    dirty: bool,
    /// Last VP matrix (f32, column-major) that was rendered into the
    /// terrain-data buffers.
    last_vp: [f32; 16],
    /// Tile set fingerprint (sorted tile IDs + generations).
    last_terrain_fingerprint: u64,
}

impl Default for TerrainDataDirtyState {
    fn default() -> Self {
        Self {
            dirty: true,
            last_vp: [0.0; 16],
            last_terrain_fingerprint: 0,
        }
    }
}

impl TerrainDataDirtyState {
    /// Check whether the terrain-data pass needs to be redrawn.
    fn needs_update(&self, vp: &glam::DMat4, terrain_meshes: &[TerrainMeshData]) -> bool {
        if self.dirty {
            return true;
        }
        let vp_f32 = vp.to_cols_array().map(|v| v as f32);
        if vp_f32 != self.last_vp {
            return true;
        }
        let fp = Self::terrain_fingerprint(terrain_meshes);
        fp != self.last_terrain_fingerprint
    }

    /// Record that the terrain-data pass was just rendered with these inputs.
    fn mark_clean(&mut self, vp: &glam::DMat4, terrain_meshes: &[TerrainMeshData]) {
        self.dirty = false;
        self.last_vp = vp.to_cols_array().map(|v| v as f32);
        self.last_terrain_fingerprint = Self::terrain_fingerprint(terrain_meshes);
    }

    fn terrain_fingerprint(terrain_meshes: &[TerrainMeshData]) -> u64 {
        let mut h: u64 = terrain_meshes.len() as u64;
        for mesh in terrain_meshes {
            h = h
                .wrapping_mul(31)
                .wrapping_add(mesh.tile.zoom as u64)
                .wrapping_mul(31)
                .wrapping_add(mesh.tile.x as u64)
                .wrapping_mul(31)
                .wrapping_add(mesh.tile.y as u64)
                .wrapping_mul(31)
                .wrapping_add(mesh.generation);
        }
        h
    }
}

#[allow(clippy::single_range_in_vec_init)]
fn diff_column_instance_ranges(
    old: &[ColumnInstanceData],
    new: &[ColumnInstanceData],
) -> Vec<std::ops::Range<usize>> {
    if old.len() != new.len() {
        return if new.is_empty() {
            Vec::new()
        } else {
            vec![(0..new.len())]
        };
    }

    let mut ranges = Vec::new();
    let mut current_start: Option<usize> = None;

    for (index, (old_item, new_item)) in old.iter().zip(new.iter()).enumerate() {
        if old_item != new_item {
            if current_start.is_none() {
                current_start = Some(index);
            }
        } else if let Some(start) = current_start.take() {
            ranges.push(start..index);
        }
    }

    if let Some(start) = current_start {
        ranges.push(start..new.len());
    }

    ranges
}

/// Cache key for tile batch buffers.  When the visible tile set and camera
/// origin haven't changed, the GPU buffers from the previous frame are
/// reused instead of being rebuilt.
#[derive(Debug, Clone, PartialEq)]
struct TileBatchCacheKey {
    /// Ordered list of (target, actual, fade_opacity_bits) tuples.
    ///
    /// `fade_opacity` is part of the key because tile vertex opacity is
    /// baked into the cached tile batch geometry.  Omitting it would allow
    /// a retained batch to freeze an in-progress fade transition until some
    /// unrelated input invalidated the cache.
    tiles: Vec<(TileId, TileId, u32)>,
    /// Camera origin quantised to avoid float drift invalidation.
    origin: [i64; 3],
    /// Active projection.
    projection: rustial_engine::CameraProjection,
}

impl TileBatchCacheKey {
    fn new(
        visible_tiles: &[VisibleTile],
        camera_origin: DVec3,
        projection: rustial_engine::CameraProjection,
    ) -> Self {
        let tiles: Vec<(TileId, TileId, u32)> = visible_tiles
            .iter()
            .map(|vt| (vt.target, vt.actual, vt.fade_opacity.to_bits()))
            .collect();
        let origin = [
            (camera_origin.x * 100.0) as i64,
            (camera_origin.y * 100.0) as i64,
            (camera_origin.z * 100.0) as i64,
        ];
        Self {
            tiles,
            origin,
            projection,
        }
    }
}

/// Cache key for vector batch buffers.  Captures a fingerprint of the
/// vector mesh data so that unchanged layers reuse their GPU buffers.
#[derive(Debug, Clone, PartialEq)]
struct VectorBatchCacheKey {
    /// Per-layer fingerprint: (vertex_count, index_count).
    layers: Vec<(usize, usize)>,
    /// Camera origin quantised to avoid float drift invalidation.
    origin: [i64; 3],
}

impl VectorBatchCacheKey {
    fn new(vector_meshes: &[VectorMeshData], camera_origin: DVec3) -> Self {
        let layers: Vec<(usize, usize)> = vector_meshes
            .iter()
            .map(|m| (m.positions.len(), m.indices.len()))
            .collect();
        let origin = [
            (camera_origin.x * 100.0) as i64,
            (camera_origin.y * 100.0) as i64,
            (camera_origin.z * 100.0) as i64,
        ];
        Self { layers, origin }
    }
}

// ---------------------------------------------------------------------------
// RenderParams
// ---------------------------------------------------------------------------

/// All inputs needed for a full render frame.
///
/// Aggregated into a single struct to keep [`WgpuMapRenderer::render_full`]'s
/// signature manageable.  Lifetimes are tied to the caller's frame data.
pub struct RenderParams<'a> {
    /// Engine map state (camera, layers, terrain manager).
    pub state: &'a MapState,
    /// WGPU device for buffer/texture/bind-group creation.
    pub device: &'a wgpu::Device,
    /// WGPU queue for `write_buffer` / `write_texture` / `submit`.
    pub queue: &'a wgpu::Queue,
    /// Colour attachment view (the surface texture for this frame).
    pub color_view: &'a wgpu::TextureView,
    /// Visible tile set for this frame (from engine's tile manager).
    pub visible_tiles: &'a [VisibleTile],
    /// Tessellated vector meshes to render (one per visible vector layer).
    pub vector_meshes: &'a [VectorMeshData],
    /// 3D model instances to render.
    pub model_instances: &'a [ModelInstance],
    /// Background / clear colour `[r, g, b, a]` in linear sRGB.
    ///
    /// Also used as the fog horizon colour.  Defaults to white if not set.
    pub clear_color: [f32; 4],
}

// ---------------------------------------------------------------------------
// WgpuMapRenderer
// ---------------------------------------------------------------------------

/// The WGPU-based map renderer.
///
/// Owns all persistent GPU resources (pipelines, uniform buffer, atlas,
/// sampler, depth texture) and provides [`render`](Self::render) /
/// [`render_full`](Self::render_full) to draw one frame.
///
/// See the [module-level documentation](self) for the full resource
/// layout and frame lifecycle.
///
/// ## Construction
///
/// ```ignore
/// let renderer = WgpuMapRenderer::new(&device, &queue, surface_format, width, height);
/// ```
///
/// ## Resize
///
/// Call [`resize`](Self::resize) whenever the surface dimensions change.
/// This recreates the depth texture.  Passing `width=0` or `height=0` is
/// clamped to 1x1 to avoid WGPU validation errors.
///
/// ## GPU batching
///
/// Tile textures are packed into shared 4096x4096 atlas pages
/// ([`TileAtlas`]).  All tiles on the same page are drawn in a single
/// batched draw call, reducing draw calls from N (one per tile) to P
/// (one per atlas page, typically 1-2).  Terrain meshes are batched
/// identically.  Vector layers are one draw call each.  Model mesh
/// GPU buffers are cached by identity fingerprint across frames.
pub struct WgpuMapRenderer {
    // -- Pipelines --------------------------------------------------------
    tile_pipeline: TilePipeline,
    terrain_pipeline: TerrainPipeline,
    terrain_data_pipeline: TerrainDataPipeline,
    hillshade_pipeline: HillshadePipeline,
    grid_scalar_pipeline: GridScalarPipeline,
    grid_extrusion_pipeline: GridExtrusionPipeline,
    column_pipeline: ColumnPipeline,
    vector_pipeline: VectorPipeline,
    fill_pipeline: FillPipeline,
    fill_pattern_pipeline: FillPatternPipeline,
    fill_extrusion_pipeline: FillExtrusionPipeline,
    line_pipeline: LinePipeline,
    line_pattern_pipeline: LinePatternPipeline,
    circle_pipeline: CirclePipeline,
    heatmap_pipeline: HeatmapPipeline,
    /// Heatmap colour-mapping pipeline (Pass 2: fullscreen ramp composite).
    heatmap_colormap_pipeline: HeatmapColormapPipeline,
    symbol_pipeline: SymbolPipeline,
    model_pipeline: ModelPipeline,
    image_overlay_pipeline: ImageOverlayPipeline,
    /// Sky / atmosphere fullscreen pipeline.
    sky_pipeline: SkyPipeline,

    // -- Shared uniform ---------------------------------------------------
    /// A single 64-byte uniform buffer holding the view-projection matrix.
    /// Shared by all four pipelines (each has its own bind group pointing
    /// to this buffer).
    uniform_buffer: wgpu::Buffer,
    /// Tile pipeline's uniform bind group (group 0).
    uniform_bind_group: wgpu::BindGroup,
    /// Terrain pipeline's uniform bind group (group 0).
    terrain_uniform_bind_group: wgpu::BindGroup,
    /// Terrain data pipeline's uniform bind group (group 0).
    terrain_data_uniform_bind_group: wgpu::BindGroup,
    /// Hillshade pipeline's uniform bind group (group 0).
    hillshade_uniform_bind_group: wgpu::BindGroup,
    /// Grid scalar pipeline's uniform bind group (group 0).
    grid_scalar_uniform_bind_group: wgpu::BindGroup,
    /// Grid extrusion pipeline's uniform bind group (group 0).
    grid_extrusion_uniform_bind_group: wgpu::BindGroup,
    /// Column pipeline's uniform bind group (group 0).
    column_uniform_bind_group: wgpu::BindGroup,
    /// Vector pipeline's uniform bind group (group 0).
    vector_uniform_bind_group: wgpu::BindGroup,
    /// Fill-extrusion pipeline's uniform bind group (group 0).
    fill_extrusion_uniform_bind_group: wgpu::BindGroup,
    /// Model pipeline's uniform bind group (group 0).
    model_uniform_bind_group: wgpu::BindGroup,
    /// Line pipeline's uniform bind group (group 0).
    line_uniform_bind_group: wgpu::BindGroup,
    /// Circle pipeline's uniform bind group (group 0).
    circle_uniform_bind_group: wgpu::BindGroup,
    /// Heatmap pipeline's uniform bind group (group 0).
    heatmap_uniform_bind_group: wgpu::BindGroup,
    /// Heatmap colormap pipeline's uniform bind group (group 0).
    heatmap_colormap_uniform_bind_group: wgpu::BindGroup,
    /// Off-screen R16Float accumulation texture for heatmap Pass 1.
    heatmap_accum_texture: wgpu::Texture,
    /// View into the accumulation texture.
    heatmap_accum_view: wgpu::TextureView,
    /// 256×1 Rgba8Unorm colour ramp texture for heatmap Pass 2.
    _heatmap_ramp_texture: wgpu::Texture,
    /// View into the colour ramp texture.
    heatmap_ramp_view: wgpu::TextureView,
    /// Heatmap colormap textures bind group (group 1: heat + ramp + sampler).
    heatmap_colormap_textures_bind_group: wgpu::BindGroup,
    /// Symbol pipeline's uniform bind group (group 0).
    symbol_uniform_bind_group: wgpu::BindGroup,
    /// Image overlay pipeline's uniform bind group (group 0).
    image_overlay_uniform_bind_group: wgpu::BindGroup,
    /// Sky pipeline's dedicated uniform buffer.
    sky_uniform_buffer: wgpu::Buffer,
    /// Sky pipeline's uniform bind group (group 0).
    sky_uniform_bind_group: wgpu::BindGroup,

    // -- Shadow map resources ---------------------------------------------
    /// Shadow depth pipeline for fill-extrusion casters.
    shadow_fill_extrusion_pipeline: crate::pipeline::shadow_pipeline::ShadowFillExtrusionPipeline,
    /// Shadow depth pipeline for model casters.
    shadow_model_pipeline: crate::pipeline::shadow_pipeline::ShadowModelPipeline,
    /// Per-cascade depth textures (one per cascade, kept alive for GPU).
    _shadow_map_textures: Vec<wgpu::Texture>,
    /// Per-cascade depth texture views.
    shadow_map_views: Vec<wgpu::TextureView>,
    /// Comparison sampler for shadow map sampling (kept alive for GPU).
    _shadow_comparison_sampler: wgpu::Sampler,
    /// Shadow params uniform buffer (2 light matrices + config).
    shadow_params_buffer: wgpu::Buffer,
    /// Shadow depth pass uniform buffer (single light VP per cascade).
    shadow_depth_uniform_buffers: Vec<wgpu::Buffer>,
    /// Shadow depth pass bind groups (group 0 for depth-only pipelines, one per cascade).
    shadow_depth_bind_groups: Vec<wgpu::BindGroup>,
    /// Shadow receiver bind group (group 1/2 for fill-extrusion / model).
    shadow_receiver_bind_group: wgpu::BindGroup,

    // -- Shared sampler ---------------------------------------------------
    /// Bilinear, clamp-to-edge sampler shared across all atlas page bind
    /// groups (tile + terrain).
    sampler: wgpu::Sampler,
    /// Filtering sampler for grid scalar ramp textures.
    grid_scalar_ramp_sampler: wgpu::Sampler,
    /// Repeat-mode sampler for fill-pattern textures.
    fill_pattern_sampler: wgpu::Sampler,

    // -- Depth ------------------------------------------------------------
    /// Depth texture view (`Depth32Float`), recreated on [`resize`](Self::resize).
    depth_view: wgpu::TextureView,
    /// Current surface width in pixels (? 1).
    width: u32,
    /// Current surface height in pixels (? 1).
    height: u32,
    /// Renderer-owned terrain depth / coordinate buffers.
    terrain_interaction_buffers: TerrainInteractionBuffers,

    // -- Atlas + page bind groups -----------------------------------------
    /// Tile texture atlas (persists across frames, evicted per-frame).
    tile_atlas: TileAtlas,
    /// Prepared hillshade texture atlas.
    hillshade_atlas: TileAtlas,
    /// Per-atlas-page bind groups for the **tile** pipeline (group 1:
    /// texture view + sampler).  Rebuilt incrementally when new pages are
    /// allocated.
    page_bind_groups: Vec<wgpu::BindGroup>,
    /// Per-atlas-page bind groups for the **terrain** pipeline (group 1).
    page_terrain_bind_groups: Vec<wgpu::BindGroup>,
    /// Per-atlas-page bind groups for the **hillshade** pipeline (group 1).
    page_hillshade_bind_groups: Vec<wgpu::BindGroup>,

    // -- Model mesh cache -------------------------------------------------
    /// Cached model mesh GPU buffers keyed by [`ModelMeshKey`] fingerprint.
    /// Avoids re-uploading identical mesh geometry every frame.
    model_mesh_cache: std::collections::HashMap<ModelMeshKey, CachedModelMesh>,
    /// Shared reusable terrain grid meshes keyed by grid resolution.
    shared_terrain_grids: std::collections::HashMap<u16, SharedTerrainGridMesh>,
    /// Cached GPU elevation textures keyed by terrain tile id.
    height_texture_cache: std::collections::HashMap<TileId, CachedHeightTexture>,
    /// Shared unit-box mesh for instanced columns.
    shared_column_mesh: Option<SharedColumnMesh>,
    /// Cached per-layer grid scalar GPU state.
    grid_scalar_overlay_cache: std::collections::HashMap<LayerId, CachedGridScalarOverlay>,
    /// Cached per-layer grid extrusion GPU state.
    grid_extrusion_overlay_cache: std::collections::HashMap<LayerId, CachedGridExtrusionOverlay>,
    /// Cached per-layer instanced column GPU state.
    column_overlay_cache: std::collections::HashMap<LayerId, CachedColumnOverlay>,
    /// Cached per-layer point-cloud GPU state.
    point_cloud_overlay_cache: std::collections::HashMap<LayerId, CachedPointCloudOverlay>,

    // -- Batch buffer caches ----------------------------------------------
    /// Cached tile batch GPU buffers from the previous frame.
    cached_tile_batches: Vec<TilePageBatches>,
    /// Cache key for the current tile batch buffers.
    tile_batch_cache_key: Option<TileBatchCacheKey>,
    /// Cached vector batch GPU buffers from the previous frame.
    cached_vector_batches: Vec<Option<VectorBatchEntry>>,
    /// Cache key for the current vector batch buffers.
    vector_batch_cache_key: Option<VectorBatchCacheKey>,
    /// Cached fill-extrusion batch GPU buffers from the previous frame.
    cached_fill_extrusion_batches: Vec<Option<FillExtrusionBatchEntry>>,
    /// Cached fill batch GPU buffers from the previous frame.
    cached_fill_batches: Vec<Option<FillBatchEntry>>,
    /// Cached fill-pattern batch GPU buffers from the previous frame.
    cached_fill_pattern_batches: Vec<Option<FillPatternBatchEntry>>,
    /// Cached line batch GPU buffers from the previous frame.
    cached_line_batches: Vec<Option<LineBatchEntry>>,
    /// Cached line-pattern batch GPU buffers from the previous frame.
    cached_line_pattern_batches: Vec<Option<LinePatternBatchEntry>>,
    /// Cached circle batch GPU buffers from the previous frame.
    cached_circle_batches: Vec<Option<CircleBatchEntry>>,
    /// Cached heatmap batch GPU buffers from the previous frame.
    cached_heatmap_batches: Vec<Option<HeatmapBatchEntry>>,
    /// Cached symbol batch GPU buffers from the previous frame.
    cached_symbol_batch: Option<SymbolBatchEntry>,
    /// GPU glyph atlas texture and view for the symbol pipeline.
    symbol_atlas_texture: Option<(wgpu::Texture, wgpu::TextureView)>,
    /// Symbol atlas bind group (group 1: texture + sampler).
    symbol_atlas_bind_group: Option<wgpu::BindGroup>,
    /// Engine-side glyph atlas used for symbol rendering.
    symbol_glyph_atlas: rustial_engine::symbols::GlyphAtlas,
    /// Glyph provider for symbol rendering (font-based or procedural).
    symbol_glyph_provider: Box<dyn rustial_engine::symbols::GlyphProvider>,

    // -- Per-tile terrain bind caches -------------------------------------
    /// Cached per-tile terrain uniform buffers and bind groups.
    terrain_tile_bind_cache: std::collections::HashMap<TerrainTileBindKey, CachedTerrainTileBind>,

    // -- Terrain-data dirty tracking --------------------------------------
    /// Dirty-tracking for the terrain-data interaction pass (MapLibre's
    /// `maybeDrawDepthAndCoords` pattern).
    terrain_data_dirty: TerrainDataDirtyState,

    // -- Model transform cache --------------------------------------------
    /// Cached model instance transform buffer + bind group.
    cached_model_transforms: Option<CachedModelTransforms>,
    /// Cached placeholder quad batch from the previous frame.
    cached_placeholder_batch: Option<VectorBatchEntry>,
    /// Cached image overlay GPU resources from the previous frame.
    cached_image_overlay_batches: Vec<CachedImageOverlayBatch>,
    /// Visualization cache activity from the last render.
    visualization_perf_stats: VisualizationPerfStats,
}

// ---------------------------------------------------------------------------
// ModelMeshKey / CachedModelMesh
// ---------------------------------------------------------------------------

/// Identity key for deduplicating model mesh GPU uploads.
///
/// Two meshes with the same `(pos_len, idx_len, fingerprint)` are assumed
/// identical.  The fingerprint is a cheap rolling hash of the first
/// position and first index -- sufficient for typical usage where distinct
/// meshes have different vertex counts.
///
/// **Limitation:** hash collisions are theoretically possible between two
/// meshes with identical lengths and coincidentally identical first
/// elements but different interiors.  A future improvement could hash the
/// full data or use an explicit user-provided mesh ID.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
struct ModelMeshKey {
    pos_len: usize,
    idx_len: usize,
    fingerprint: u64,
}

impl ModelMeshKey {
    fn from_mesh(mesh: &rustial_engine::ModelMesh) -> Self {
        let mut fingerprint: u64 = mesh.positions.len() as u64;
        if let Some(first) = mesh.positions.first() {
            fingerprint = fingerprint
                .wrapping_mul(31)
                .wrapping_add(first[0].to_bits() as u64)
                .wrapping_mul(31)
                .wrapping_add(first[1].to_bits() as u64)
                .wrapping_mul(31)
                .wrapping_add(first[2].to_bits() as u64);
        }
        if let Some(&first_idx) = mesh.indices.first() {
            fingerprint = fingerprint.wrapping_mul(31).wrapping_add(first_idx as u64);
        }
        Self {
            pos_len: mesh.positions.len(),
            idx_len: mesh.indices.len(),
            fingerprint,
        }
    }
}

/// A model mesh whose vertex + index buffers have been uploaded to the GPU.
struct CachedModelMesh {
    vertex_buffer: wgpu::Buffer,
    index_buffer: wgpu::Buffer,
    index_count: u32,
}

/// Cached model instance transform buffer and bind group.
///
/// Avoids re-creating the GPU buffer and bind group every frame when
/// the model instance list and camera origin haven't changed.
struct CachedModelTransforms {
    #[allow(dead_code)]
    buffer: wgpu::Buffer,
    bind_group: wgpu::BindGroup,
    /// Stride in bytes between consecutive instance transforms.
    stride: usize,
    /// Instance count at the time this was created.
    instance_count: usize,
    /// Rolling fingerprint of (instance positions + rotations + scales
    /// + camera origin).
    fingerprint: u64,
}

/// Cached GPU resources for a single image overlay.
struct CachedImageOverlayBatch {
    vertex_buffer: wgpu::Buffer,
    index_buffer: wgpu::Buffer,
    texture: wgpu::Texture,
    #[allow(dead_code)]
    texture_view: wgpu::TextureView,
    texture_bind_group: wgpu::BindGroup,
    /// Layer id that produced this overlay (for cache key matching).
    layer_id: rustial_engine::LayerId,
    /// Texture dimensions `(width, height)` for reuse checks.
    tex_dimensions: (u32, u32),
    /// Data pointer identity for fast same-frame skip.
    data_arc_ptr: usize,
}

// ---------------------------------------------------------------------------
// impl WgpuMapRenderer
// ---------------------------------------------------------------------------

impl WgpuMapRenderer {
    /// Create a new renderer.
    ///
    /// # Arguments
    ///
    /// * `device` -- WGPU device.
    /// * `_queue` -- WGPU queue (reserved for future lazy init; unused today).
    /// * `format` -- Colour target format (must match the surface's preferred format).
    /// * `width`  -- Initial surface width in pixels.
    /// * `height` -- Initial surface height in pixels.
    pub fn new(
        device: &wgpu::Device,
        _queue: &wgpu::Queue,
        format: wgpu::TextureFormat,
        width: u32,
        height: u32,
    ) -> Self {
        let tile_pipeline = TilePipeline::new(device, format);
        let terrain_pipeline =
            TerrainPipeline::new(device, format, &tile_pipeline.uniform_bind_group_layout);
        let terrain_data_pipeline = TerrainDataPipeline::new(device);
        let hillshade_pipeline = HillshadePipeline::new(device, format);
        let grid_scalar_pipeline = GridScalarPipeline::new(device, format);
        let grid_extrusion_pipeline = GridExtrusionPipeline::new(device, format);
        let column_pipeline = ColumnPipeline::new(device, format);
        let vector_pipeline = VectorPipeline::new(device, format);
        let fill_pipeline = FillPipeline::new(device, format);
        let fill_pattern_pipeline = FillPatternPipeline::new(device, format);
        let fill_extrusion_pipeline = FillExtrusionPipeline::new(device, format);
        let line_pipeline = LinePipeline::new(device, format);
        let line_pattern_pipeline = LinePatternPipeline::new(device, format);
        let circle_pipeline = CirclePipeline::new(device, format);
        let heatmap_pipeline = HeatmapPipeline::new(device);
        let heatmap_colormap_pipeline = HeatmapColormapPipeline::new(device, format);
        let symbol_pipeline = SymbolPipeline::new(device, format);
        let model_pipeline = ModelPipeline::new(device, format);
        let image_overlay_pipeline = ImageOverlayPipeline::new(device, format);
        let sky_pipeline = SkyPipeline::new(device, format);

        // Shared uniform buffer (view-projection + fog parameters).
        let uniform_data = ViewProjUniform::from_dmat4(&glam::DMat4::IDENTITY);
        let uniform_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
            label: Some("rustial_uniform_buf"),
            contents: bytemuck::bytes_of(&uniform_data),
            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
        });

        // Four bind groups pointing to the same buffer -- one per pipeline.
        // Each pipeline may have a different `BindGroupLayout` (even though
        // the layout *happens* to be identical today) so we create separate
        // bind groups to stay correct if layouts diverge.
        let uniform_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_uniform_bg"),
            layout: &tile_pipeline.uniform_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: uniform_buffer.as_entire_binding(),
            }],
        });

        let column_uniform_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_column_uniform_bg"),
            layout: &column_pipeline.uniform_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: uniform_buffer.as_entire_binding(),
            }],
        });

        let grid_extrusion_uniform_bind_group =
            device.create_bind_group(&wgpu::BindGroupDescriptor {
                label: Some("rustial_grid_extrusion_uniform_bg"),
                layout: &grid_extrusion_pipeline.uniform_bind_group_layout,
                entries: &[wgpu::BindGroupEntry {
                    binding: 0,
                    resource: uniform_buffer.as_entire_binding(),
                }],
            });

        let terrain_uniform_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_terrain_uniform_bg"),
            layout: &terrain_pipeline.uniform_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: uniform_buffer.as_entire_binding(),
            }],
        });

        let hillshade_uniform_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_hillshade_uniform_bg"),
            layout: &hillshade_pipeline.uniform_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: uniform_buffer.as_entire_binding(),
            }],
        });

        let grid_scalar_uniform_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_grid_scalar_uniform_bg"),
            layout: &grid_scalar_pipeline.uniform_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: uniform_buffer.as_entire_binding(),
            }],
        });

        let vector_uniform_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_vector_uniform_bg"),
            layout: &vector_pipeline.uniform_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: uniform_buffer.as_entire_binding(),
            }],
        });

        let fill_extrusion_uniform_bind_group =
            device.create_bind_group(&wgpu::BindGroupDescriptor {
                label: Some("rustial_fill_extrusion_uniform_bg"),
                layout: &fill_extrusion_pipeline.uniform_bind_group_layout,
                entries: &[wgpu::BindGroupEntry {
                    binding: 0,
                    resource: uniform_buffer.as_entire_binding(),
                }],
            });

        let model_uniform_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_model_uniform_bg"),
            layout: &model_pipeline.uniform_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: uniform_buffer.as_entire_binding(),
            }],
        });

        let line_uniform_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_line_uniform_bg"),
            layout: &line_pipeline.uniform_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: uniform_buffer.as_entire_binding(),
            }],
        });

        let circle_uniform_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_circle_uniform_bg"),
            layout: &circle_pipeline.uniform_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: uniform_buffer.as_entire_binding(),
            }],
        });

        let heatmap_uniform_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_heatmap_uniform_bg"),
            layout: &heatmap_pipeline.uniform_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: uniform_buffer.as_entire_binding(),
            }],
        });

        let heatmap_colormap_uniform_bind_group =
            device.create_bind_group(&wgpu::BindGroupDescriptor {
                label: Some("rustial_heatmap_colormap_uniform_bg"),
                layout: &heatmap_colormap_pipeline.uniform_bind_group_layout,
                entries: &[wgpu::BindGroupEntry {
                    binding: 0,
                    resource: uniform_buffer.as_entire_binding(),
                }],
            });

        let symbol_uniform_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_symbol_uniform_bg"),
            layout: &symbol_pipeline.uniform_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: uniform_buffer.as_entire_binding(),
            }],
        });

        let image_overlay_uniform_bind_group =
            device.create_bind_group(&wgpu::BindGroupDescriptor {
                label: Some("rustial_image_overlay_uniform_bg"),
                layout: &image_overlay_pipeline.uniform_bind_group_layout,
                entries: &[wgpu::BindGroupEntry {
                    binding: 0,
                    resource: uniform_buffer.as_entire_binding(),
                }],
            });

        // Sky pipeline has its own dedicated uniform buffer (different layout).
        let sky_uniform_data = SkyUniform::default();
        let sky_uniform_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
            label: Some("rustial_sky_uniform_buf"),
            contents: bytemuck::bytes_of(&sky_uniform_data),
            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
        });
        let sky_uniform_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_sky_uniform_bg"),
            layout: &sky_pipeline.uniform_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: sky_uniform_buffer.as_entire_binding(),
            }],
        });

        let terrain_data_uniform_bind_group =
            device.create_bind_group(&wgpu::BindGroupDescriptor {
                label: Some("rustial_terrain_data_uniform_bg"),
                layout: &terrain_data_pipeline.uniform_bind_group_layout,
                entries: &[wgpu::BindGroupEntry {
                    binding: 0,
                    resource: uniform_buffer.as_entire_binding(),
                }],
            });

        let sampler = device.create_sampler(&wgpu::SamplerDescriptor {
            label: Some("rustial_sampler"),
            address_mode_u: wgpu::AddressMode::ClampToEdge,
            address_mode_v: wgpu::AddressMode::ClampToEdge,
            mag_filter: wgpu::FilterMode::Linear,
            min_filter: wgpu::FilterMode::Linear,
            mipmap_filter: wgpu::FilterMode::Linear,
            anisotropy_clamp: 16,
            ..Default::default()
        });

        let grid_scalar_ramp_sampler = device.create_sampler(&wgpu::SamplerDescriptor {
            label: Some("rustial_grid_scalar_ramp_sampler"),
            address_mode_u: wgpu::AddressMode::ClampToEdge,
            address_mode_v: wgpu::AddressMode::ClampToEdge,
            mag_filter: wgpu::FilterMode::Linear,
            min_filter: wgpu::FilterMode::Linear,
            mipmap_filter: wgpu::FilterMode::Nearest,
            ..Default::default()
        });

        let fill_pattern_sampler = device.create_sampler(&wgpu::SamplerDescriptor {
            label: Some("rustial_fill_pattern_sampler"),
            address_mode_u: wgpu::AddressMode::Repeat,
            address_mode_v: wgpu::AddressMode::Repeat,
            mag_filter: wgpu::FilterMode::Linear,
            min_filter: wgpu::FilterMode::Linear,
            mipmap_filter: wgpu::FilterMode::Linear,
            ..Default::default()
        });

        // Clamp to at least 1x1 so the depth texture is always valid.
        let w = width.max(1);
        let h = height.max(1);
        let depth_view = create_depth_texture(device, w, h);
        let terrain_interaction_buffers = TerrainInteractionBuffers::new(device, w, h);

        // -- Heatmap off-screen resources ---------------------------------
        let (heatmap_accum_texture, heatmap_accum_view) =
            create_heatmap_accum_texture(device, w, h);
        let heatmap_ramp_texture = create_default_heatmap_ramp_texture(device, _queue);
        let heatmap_ramp_view =
            heatmap_ramp_texture.create_view(&wgpu::TextureViewDescriptor::default());
        let heatmap_colormap_textures_bind_group = create_heatmap_colormap_bind_group(
            device,
            &heatmap_colormap_pipeline.textures_bind_group_layout,
            &heatmap_accum_view,
            &heatmap_ramp_view,
            &sampler,
        );

        // -- Shadow map resources -----------------------------------------
        let shadow_fill_extrusion_pipeline =
            crate::pipeline::shadow_pipeline::ShadowFillExtrusionPipeline::new(device);
        let shadow_model_pipeline =
            crate::pipeline::shadow_pipeline::ShadowModelPipeline::new(device);

        let shadow_map_resolution = 2048u32;
        let (shadow_map_textures, shadow_map_views) =
            create_shadow_map_textures(device, shadow_map_resolution, 2);

        let shadow_comparison_sampler = device.create_sampler(&wgpu::SamplerDescriptor {
            label: Some("rustial_shadow_comparison_sampler"),
            compare: Some(wgpu::CompareFunction::LessEqual),
            mag_filter: wgpu::FilterMode::Linear,
            min_filter: wgpu::FilterMode::Linear,
            ..Default::default()
        });

        let shadow_params_buffer = device.create_buffer(&wgpu::BufferDescriptor {
            label: Some("rustial_shadow_params_buffer"),
            size: 160, // 2 × mat4 (128) + 2 × vec4 (32) = 160
            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
            mapped_at_creation: false,
        });

        let shadow_depth_uniform_buffers: Vec<wgpu::Buffer> = (0..2)
            .map(|i| {
                device.create_buffer(&wgpu::BufferDescriptor {
                    label: Some(&format!("rustial_shadow_depth_uniform_{i}")),
                    size: 64, // mat4x4<f32>
                    usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
                    mapped_at_creation: false,
                })
            })
            .collect();

        let shadow_depth_bind_groups: Vec<wgpu::BindGroup> = shadow_depth_uniform_buffers
            .iter()
            .enumerate()
            .map(|(i, buf)| {
                device.create_bind_group(&wgpu::BindGroupDescriptor {
                    label: Some(&format!("rustial_shadow_depth_bg_{i}")),
                    layout: &shadow_fill_extrusion_pipeline.uniform_bind_group_layout,
                    entries: &[wgpu::BindGroupEntry {
                        binding: 0,
                        resource: buf.as_entire_binding(),
                    }],
                })
            })
            .collect();

        let shadow_receiver_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("rustial_shadow_receiver_bg"),
            layout: &fill_extrusion_pipeline.shadow_bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: shadow_params_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: wgpu::BindingResource::TextureView(&shadow_map_views[0]),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: wgpu::BindingResource::TextureView(&shadow_map_views[1]),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: wgpu::BindingResource::Sampler(&shadow_comparison_sampler),
                },
            ],
        });

        Self {
            tile_pipeline,
            terrain_pipeline,
            terrain_data_pipeline,
            hillshade_pipeline,
            grid_scalar_pipeline,
            grid_extrusion_pipeline,
            column_pipeline,
            vector_pipeline,
            fill_pipeline,
            fill_pattern_pipeline,
            fill_extrusion_pipeline,
            line_pipeline,
            line_pattern_pipeline,
            circle_pipeline,
            heatmap_pipeline,
            heatmap_colormap_pipeline,
            symbol_pipeline,
            model_pipeline,
            image_overlay_pipeline,
            sky_pipeline,
            uniform_buffer,
            uniform_bind_group,
            terrain_uniform_bind_group,
            terrain_data_uniform_bind_group,
            hillshade_uniform_bind_group,
            grid_scalar_uniform_bind_group,
            grid_extrusion_uniform_bind_group,
            column_uniform_bind_group,
            vector_uniform_bind_group,
            fill_extrusion_uniform_bind_group,
            model_uniform_bind_group,
            line_uniform_bind_group,
            circle_uniform_bind_group,
            heatmap_uniform_bind_group,
            heatmap_colormap_uniform_bind_group,
            heatmap_accum_texture,
            heatmap_accum_view,
            _heatmap_ramp_texture: heatmap_ramp_texture,
            heatmap_ramp_view,
            heatmap_colormap_textures_bind_group,
            symbol_uniform_bind_group,
            image_overlay_uniform_bind_group,
            sky_uniform_buffer,
            sky_uniform_bind_group,
            shadow_fill_extrusion_pipeline,
            shadow_model_pipeline,
            _shadow_map_textures: shadow_map_textures,
            shadow_map_views,
            _shadow_comparison_sampler: shadow_comparison_sampler,
            shadow_params_buffer,
            shadow_depth_uniform_buffers,
            shadow_depth_bind_groups,
            shadow_receiver_bind_group,
            sampler,
            grid_scalar_ramp_sampler,
            fill_pattern_sampler,
            depth_view,
            width: w,
            height: h,
            terrain_interaction_buffers,
            tile_atlas: TileAtlas::new(),
            hillshade_atlas: TileAtlas::new(),
            page_bind_groups: Vec::new(),
            page_terrain_bind_groups: Vec::new(),
            page_hillshade_bind_groups: Vec::new(),
            model_mesh_cache: std::collections::HashMap::new(),
            shared_terrain_grids: std::collections::HashMap::new(),
            height_texture_cache: std::collections::HashMap::new(),
            shared_column_mesh: None,
            grid_scalar_overlay_cache: std::collections::HashMap::new(),
            grid_extrusion_overlay_cache: std::collections::HashMap::new(),
            column_overlay_cache: std::collections::HashMap::new(),
            point_cloud_overlay_cache: std::collections::HashMap::new(),
            cached_tile_batches: Vec::new(),
            tile_batch_cache_key: None,
            cached_vector_batches: Vec::new(),
            vector_batch_cache_key: None,
            cached_fill_extrusion_batches: Vec::new(),
            cached_fill_batches: Vec::new(),
            cached_fill_pattern_batches: Vec::new(),
            cached_line_batches: Vec::new(),
            cached_line_pattern_batches: Vec::new(),
            cached_circle_batches: Vec::new(),
            cached_heatmap_batches: Vec::new(),
            cached_symbol_batch: None,
            symbol_atlas_texture: None,
            symbol_atlas_bind_group: None,
            symbol_glyph_atlas: rustial_engine::symbols::GlyphAtlas::new(),
            symbol_glyph_provider: Box::new(rustial_engine::symbols::ProceduralGlyphProvider::new()),
            terrain_tile_bind_cache: std::collections::HashMap::new(),
            terrain_data_dirty: TerrainDataDirtyState::default(),
            cached_model_transforms: None,
            cached_placeholder_batch: None,
            cached_image_overlay_batches: Vec::new(),
            visualization_perf_stats: VisualizationPerfStats::default(),
        }
    }

    // -- Surface management -----------------------------------------------

    /// Notify the renderer that the surface was resized.
    ///
    /// Recreates the depth texture.  Dimensions are clamped to at least
    /// 1x1 -- passing `0` is safe and produces a 1-pixel texture.
    pub fn resize(&mut self, device: &wgpu::Device, width: u32, height: u32) {
        self.width = width.max(1);
        self.height = height.max(1);
        self.depth_view = create_depth_texture(device, self.width, self.height);
        self.terrain_interaction_buffers
            .resize(device, self.width, self.height);
        self.terrain_data_dirty.dirty = true;

        // Recreate heatmap accumulation texture at new size and rebuild the
        // colour-map bind group that references it.
        let (tex, view) = create_heatmap_accum_texture(device, self.width, self.height);
        self.heatmap_accum_texture = tex;
        self.heatmap_accum_view = view;
        self.heatmap_colormap_textures_bind_group = create_heatmap_colormap_bind_group(
            device,
            &self.heatmap_colormap_pipeline.textures_bind_group_layout,
            &self.heatmap_accum_view,
            &self.heatmap_ramp_view,
            &self.sampler,
        );
    }

    /// Replace the glyph provider used for symbol text rendering.
    ///
    /// By default the renderer uses [`ProceduralGlyphProvider`] which
    /// produces placeholder glyphs.  Pass a
    /// [`ShapedGlyphProvider`](rustial_engine::symbols::text_shaper::ShapedGlyphProvider)
    /// (when the `text-shaping` feature is enabled) to render real
    /// font-based SDF text.
    pub fn set_glyph_provider(
        &mut self,
        provider: Box<dyn rustial_engine::symbols::GlyphProvider>,
    ) {
        self.symbol_glyph_provider = provider;
    }

    // -- Tile upload ------------------------------------------------------

    /// Enqueue a decoded tile image for deferred GPU upload.
    ///
    /// If the tile is already present this is a no-op.  Otherwise the slot
    /// is allocated and a deferred upload is enqueued.  The actual GPU
    /// `write_texture` happens when [`flush_atlas_uploads`] is called
    /// during the frame.  Page bind groups are rebuilt if a new atlas page
    /// was created.
    pub fn upload_tile(&mut self, device: &wgpu::Device, tile_id: TileId, image: &DecodedImage) {
        if self.tile_atlas.contains(&tile_id) {
            return;
        }

        if let Err(err) = image.validate_rgba8() {
            log::warn!(
                "wgpu upload_tile: skipping invalid tile {:?}: {}",
                tile_id,
                err
            );
            return;
        }

        self.tile_atlas.insert(device, tile_id, image);
        self.tile_batch_cache_key = None;

        // Rebuild page bind groups if new pages were created.
        self.rebuild_page_bind_groups(device);
    }

    /// Enqueue a prepared hillshade texture for deferred GPU upload.
    pub fn upload_hillshade(
        &mut self,
        device: &wgpu::Device,
        tile_id: TileId,
        image: &DecodedImage,
    ) {
        if self.hillshade_atlas.contains(&tile_id) {
            return;
        }
        if let Err(err) = image.validate_rgba8() {
            log::warn!(
                "wgpu upload_hillshade: skipping invalid tile {:?}: {}",
                tile_id,
                err
            );
            return;
        }
        self.hillshade_atlas.insert(device, tile_id, image);
        self.rebuild_page_bind_groups(device);
    }

    /// Flush all pending atlas texture uploads to the GPU.
    ///
    /// Writes only the affected slot pixel rectangles (partial writes)
    /// rather than re-uploading full atlas pages.  Call once per frame
    /// after all [`upload_tile`](Self::upload_tile) /
    /// [`upload_hillshade`](Self::upload_hillshade) calls and before
    /// building batched geometry.
    pub fn flush_atlas_uploads(&mut self, queue: &wgpu::Queue) {
        self.tile_atlas.flush_uploads(queue);
        self.hillshade_atlas.flush_uploads(queue);
    }

    fn get_or_create_shared_column_mesh(&mut self, device: &wgpu::Device) -> &SharedColumnMesh {
        if self.shared_column_mesh.is_none() {
            let (vertices, indices) = build_unit_column_mesh();
            let vertex_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("column_unit_box_vb"),
                contents: bytemuck::cast_slice(&vertices),
                usage: wgpu::BufferUsages::VERTEX,
            });
            let index_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("column_unit_box_ib"),
                contents: bytemuck::cast_slice(&indices),
                usage: wgpu::BufferUsages::INDEX,
            });
            self.shared_column_mesh = Some(SharedColumnMesh {
                vertex_buffer,
                index_buffer,
                index_count: indices.len() as u32,
            });
        }
        self.shared_column_mesh.as_ref().expect("column mesh")
    }

    fn get_or_create_grid_scalar_overlay(
        &mut self,
        device: &wgpu::Device,
        queue: &wgpu::Queue,
        overlay: &VisualizationOverlay,
        state: &MapState,
        scene_origin: DVec3,
        terrain_fingerprint: u64,
    ) -> Option<()> {
        let VisualizationOverlay::GridScalar {
            layer_id,
            grid,
            field,
            ramp,
        } = overlay
        else {
            return None;
        };

        let origin_key = [
            (scene_origin.x * 100.0) as i64,
            (scene_origin.y * 100.0) as i64,
            (scene_origin.z * 100.0) as i64,
        ];
        let ramp_fingerprint = grid_scalar_ramp_fingerprint(ramp);
        let grid_fingerprint = grid_extrusion_grid_fingerprint(grid);
        let projection = state.camera().projection();
        let (vertices, indices) = build_grid_scalar_geometry(grid, state, scene_origin);

        let recreate = if let Some(cached) = self.grid_scalar_overlay_cache.get(layer_id) {
            cached.generation != field.generation
                || cached.ramp_fingerprint != ramp_fingerprint
                || cached.grid_fingerprint != grid_fingerprint
                || cached.projection != projection
                || cached.index_count as usize != indices.len()
                || cached.vertex_count != vertices.len()
        } else {
            true
        };

        if recreate {
            self.visualization_perf_stats.grid_scalar_rebuilds += 1;
            let vertex_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some(&format!("grid_scalar_vb_{layer_id}")),
                contents: bytemuck::cast_slice(&vertices),
                usage: wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
            });
            let index_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some(&format!("grid_scalar_ib_{layer_id}")),
                contents: bytemuck::cast_slice(&indices),
                usage: wgpu::BufferUsages::INDEX | wgpu::BufferUsages::COPY_DST,
            });
            let scalar_texture = create_grid_scalar_texture(device, queue, field);
            let scalar_view = scalar_texture.create_view(&wgpu::TextureViewDescriptor::default());
            let ramp_texture = create_grid_scalar_ramp_texture(device, queue, ramp);
            let ramp_view = ramp_texture.create_view(&wgpu::TextureViewDescriptor::default());
            let uniform = build_grid_scalar_uniform(grid, field, state, scene_origin, 1.0);
            let uniform_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some(&format!("grid_scalar_uniform_{layer_id}")),
                contents: bytemuck::bytes_of(&uniform),
                usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
            });
            let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
                label: Some(&format!("grid_scalar_bg_{layer_id}")),
                layout: &self.grid_scalar_pipeline.overlay_bind_group_layout,
                entries: &[
                    wgpu::BindGroupEntry {
                        binding: 0,
                        resource: uniform_buffer.as_entire_binding(),
                    },
                    wgpu::BindGroupEntry {
                        binding: 1,
                        resource: wgpu::BindingResource::TextureView(&scalar_view),
                    },
                    wgpu::BindGroupEntry {
                        binding: 2,
                        resource: wgpu::BindingResource::TextureView(&ramp_view),
                    },
                    wgpu::BindGroupEntry {
                        binding: 3,
                        resource: wgpu::BindingResource::Sampler(&self.grid_scalar_ramp_sampler),
                    },
                ],
            });
            self.grid_scalar_overlay_cache.insert(
                *layer_id,
                CachedGridScalarOverlay {
                    vertex_buffer,
                    index_buffer,
                    index_count: indices.len() as u32,
                    vertex_count: vertices.len(),
                    uniform_buffer,
                    bind_group,
                    scalar_texture,
                    ramp_texture,
                    generation: field.generation,
                    value_generation: field.value_generation,
                    ramp_fingerprint,
                    grid_fingerprint,
                    terrain_fingerprint,
                    projection,
                    origin_key,
                },
            );
            return Some(());
        }

        if let Some(cached) = self.grid_scalar_overlay_cache.get_mut(layer_id) {
            let uniform = build_grid_scalar_uniform(grid, field, state, scene_origin, 1.0);
            if cached.value_generation != field.value_generation {
                self.visualization_perf_stats.grid_scalar_value_updates += 1;
                write_grid_scalar_texture(queue, &cached.scalar_texture, field);
                cached.value_generation = field.value_generation;
                queue.write_buffer(&cached.uniform_buffer, 0, bytemuck::bytes_of(&uniform));
            }
            if cached.origin_key != origin_key || cached.terrain_fingerprint != terrain_fingerprint
            {
                queue.write_buffer(
                    &cached.vertex_buffer,
                    0,
                    bytemuck::cast_slice::<GridScalarVertex, u8>(&vertices),
                );
                queue.write_buffer(&cached.uniform_buffer, 0, bytemuck::bytes_of(&uniform));
                cached.origin_key = origin_key;
                cached.terrain_fingerprint = terrain_fingerprint;
            }
        }

        Some(())
    }

    fn get_or_create_point_cloud_overlay(
        &mut self,
        device: &wgpu::Device,
        queue: &wgpu::Queue,
        overlay: &VisualizationOverlay,
        state: &MapState,
        scene_origin: DVec3,
    ) -> Option<()> {
        let VisualizationOverlay::Points {
            layer_id,
            points,
            ramp,
        } = overlay
        else {
            return None;
        };

        let origin_key = [
            (scene_origin.x * 100.0) as i64,
            (scene_origin.y * 100.0) as i64,
            (scene_origin.z * 100.0) as i64,
        ];
        let points_fingerprint = point_set_fingerprint(points);
        let ramp_fingerprint = grid_scalar_ramp_fingerprint(ramp);
        let instances = build_point_instances(points, ramp, state, scene_origin);

        let needs_rebuild = if let Some(cached) = self.point_cloud_overlay_cache.get(layer_id) {
            cached.generation != points.generation
                || cached.ramp_fingerprint != ramp_fingerprint
                || cached.instance_count as usize != instances.len()
        } else {
            true
        };

        if needs_rebuild {
            self.visualization_perf_stats.point_cloud_rebuilds += 1;
            let instance_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some(&format!("point_cloud_instances_{layer_id}")),
                contents: bytemuck::cast_slice(&instances),
                usage: wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
            });
            self.point_cloud_overlay_cache.insert(
                *layer_id,
                CachedPointCloudOverlay {
                    instance_buffer,
                    instance_count: instances.len() as u32,
                    generation: points.generation,
                    origin_key,
                    points_fingerprint,
                    ramp_fingerprint,
                    instance_data: instances,
                },
            );
            return Some(());
        }

        if let Some(cached) = self.point_cloud_overlay_cache.get_mut(layer_id) {
            let ranges = diff_column_instance_ranges(&cached.instance_data, &instances);
            if !ranges.is_empty() {
                self.visualization_perf_stats.point_cloud_partial_writes += 1;
                self.visualization_perf_stats
                    .point_cloud_partial_write_ranges += ranges.len() as u32;
            }
            for range in ranges {
                let start = range.start;
                let end = range.end;
                let byte_offset =
                    (start * std::mem::size_of::<ColumnInstanceData>()) as wgpu::BufferAddress;
                queue.write_buffer(
                    &cached.instance_buffer,
                    byte_offset,
                    bytemuck::cast_slice::<ColumnInstanceData, u8>(&instances[start..end]),
                );
            }
            cached.instance_data = instances;
            cached.origin_key = origin_key;
            cached.points_fingerprint = points_fingerprint;
        }

        Some(())
    }

    fn get_or_create_grid_extrusion_overlay(
        &mut self,
        device: &wgpu::Device,
        queue: &wgpu::Queue,
        overlay: &VisualizationOverlay,
        state: &MapState,
        scene_origin: DVec3,
        terrain_fingerprint: u64,
    ) -> Option<()> {
        let VisualizationOverlay::GridExtrusion {
            layer_id,
            grid,
            field,
            ramp,
            params,
        } = overlay
        else {
            return None;
        };

        let origin_key = [
            (scene_origin.x * 100.0) as i64,
            (scene_origin.y * 100.0) as i64,
            (scene_origin.z * 100.0) as i64,
        ];
        let grid_fingerprint = grid_extrusion_grid_fingerprint(grid);
        let params_fingerprint = grid_extrusion_params_fingerprint(params);
        let ramp_fingerprint = grid_scalar_ramp_fingerprint(ramp);

        let (vertices, indices) =
            build_grid_extrusion_geometry(grid, field, ramp, params, state, scene_origin);

        let needs_rebuild = if let Some(cached) = self.grid_extrusion_overlay_cache.get(layer_id) {
            cached.generation != field.generation
                || cached.grid_fingerprint != grid_fingerprint
                || cached.params_fingerprint != params_fingerprint
                || cached.ramp_fingerprint != ramp_fingerprint
                || cached.terrain_fingerprint != terrain_fingerprint
                || cached.index_count as usize != indices.len()
                || cached.vertex_count != vertices.len()
        } else {
            true
        };

        if needs_rebuild {
            self.visualization_perf_stats.grid_extrusion_rebuilds += 1;
            let vertex_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some(&format!("grid_extrusion_vb_{layer_id}")),
                contents: bytemuck::cast_slice(&vertices),
                usage: wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
            });
            let index_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some(&format!("grid_extrusion_ib_{layer_id}")),
                contents: bytemuck::cast_slice(&indices),
                usage: wgpu::BufferUsages::INDEX | wgpu::BufferUsages::COPY_DST,
            });

            self.grid_extrusion_overlay_cache.insert(
                *layer_id,
                CachedGridExtrusionOverlay {
                    vertex_buffer,
                    index_buffer,
                    index_count: indices.len() as u32,
                    vertex_count: vertices.len(),
                    generation: field.generation,
                    value_generation: field.value_generation,
                    origin_key,
                    grid_fingerprint,
                    params_fingerprint,
                    ramp_fingerprint,
                    terrain_fingerprint,
                },
            );
            return Some(());
        }

        if let Some(cached) = self.grid_extrusion_overlay_cache.get_mut(layer_id) {
            if cached.value_generation != field.value_generation || cached.origin_key != origin_key
            {
                self.visualization_perf_stats.grid_extrusion_value_updates += 1;
                let vertex_bytes = bytemuck::cast_slice::<GridExtrusionVertex, u8>(&vertices);
                queue.write_buffer(&cached.vertex_buffer, 0, vertex_bytes);
            }
            cached.value_generation = field.value_generation;
            cached.origin_key = origin_key;
            cached.terrain_fingerprint = terrain_fingerprint;
        }
        Some(())
    }

    fn get_or_create_column_overlay(
        &mut self,
        device: &wgpu::Device,
        queue: &wgpu::Queue,
        overlay: &VisualizationOverlay,
        state: &MapState,
        scene_origin: DVec3,
    ) -> Option<()> {
        let VisualizationOverlay::Columns {
            layer_id,
            columns,
            ramp,
        } = overlay
        else {
            return None;
        };

        let origin_key = [
            (scene_origin.x * 100.0) as i64,
            (scene_origin.y * 100.0) as i64,
            (scene_origin.z * 100.0) as i64,
        ];
        let columns_fingerprint = column_set_fingerprint(columns);
        let ramp_fingerprint = grid_scalar_ramp_fingerprint(ramp);
        let instances = build_column_instances(columns, ramp, state, scene_origin);

        let needs_rebuild = if let Some(cached) = self.column_overlay_cache.get(layer_id) {
            cached.generation != columns.generation
                || cached.ramp_fingerprint != ramp_fingerprint
                || cached.instance_count as usize != instances.len()
        } else {
            true
        };

        if needs_rebuild {
            self.visualization_perf_stats.column_rebuilds += 1;
            let instance_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some(&format!("column_instances_{layer_id}")),
                contents: bytemuck::cast_slice(&instances),
                usage: wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
            });
            self.column_overlay_cache.insert(
                *layer_id,
                CachedColumnOverlay {
                    instance_buffer,
                    instance_count: instances.len() as u32,
                    generation: columns.generation,
                    origin_key,
                    columns_fingerprint,
                    ramp_fingerprint,
                    instance_data: instances,
                },
            );
            return Some(());
        }

        if let Some(cached) = self.column_overlay_cache.get_mut(layer_id) {
            let ranges = diff_column_instance_ranges(&cached.instance_data, &instances);
            if !ranges.is_empty() {
                self.visualization_perf_stats.column_partial_writes += 1;
                self.visualization_perf_stats.column_partial_write_ranges += ranges.len() as u32;
            }
            for range in ranges {
                let start = range.start;
                let end = range.end;
                let byte_offset =
                    (start * std::mem::size_of::<ColumnInstanceData>()) as wgpu::BufferAddress;
                queue.write_buffer(
                    &cached.instance_buffer,
                    byte_offset,
                    bytemuck::cast_slice::<ColumnInstanceData, u8>(&instances[start..end]),
                );
            }
            cached.instance_data = instances;
            cached.origin_key = origin_key;
            cached.columns_fingerprint = columns_fingerprint;
        }

        Some(())
    }

    // -- Render entry points ----------------------------------------------

    /// Render one frame of the map (tiles only, no vectors or models).
    ///
    /// Convenience wrapper around [`render_full`](Self::render_full) that
    /// passes empty slices for `vector_meshes` and `model_instances`.
    pub fn render(
        &mut self,
        state: &MapState,
        device: &wgpu::Device,
        queue: &wgpu::Queue,
        color_view: &wgpu::TextureView,
        visible_tiles: &[VisibleTile],
    ) {
        let clear_color = state.computed_fog().clear_color;
        self.render_full(&RenderParams {
            state,
            device,
            queue,
            color_view,
            visible_tiles,
            vector_meshes: &[],
            model_instances: &[],
            clear_color,
        });
    }

    /// Render one full frame: tiles (or terrain), vectors, and models.
    ///
    /// See the [module-level frame lifecycle](self) for the step-by-step
    /// breakdown.
    ///
    /// ## Batching strategy
    ///
    /// 1. **Tiles / terrain** -- all quads / meshes sharing the same atlas
    ///    page are merged into a single vertex + index buffer and drawn
    ///    with one `draw_indexed` call per page.
    /// 2. **Vectors** -- each vector layer produces one draw call (already
    ///    pre-merged by the engine tessellator).
    /// 3. **Models** -- mesh GPU buffers are cached by identity fingerprint;
    ///    per-instance transform uniform + bind group are allocated
    ///    per-frame (future: dynamic UBO).
    pub fn render_full(&mut self, params: &RenderParams<'_>) {
        self.visualization_perf_stats = VisualizationPerfStats::default();
        // ?? 1. Camera-relative view-projection ??????????????????????????
        let scene_camera_origin = params.state.scene_world_origin();
        let frame = params.state.frame_output();
        let visualization = &frame.visualization;
        let view = params.state.camera().view_matrix(DVec3::ZERO);
        let proj = params.state.camera().projection_matrix();
        let vp = proj * view;

        // Fog: read pre-computed fog from the engine (centralised,
        // replaces the duplicated fog math that was previously inline here).
        let cam = params.state.camera();
        let eye = cam.eye_offset();
        let fog = params.state.computed_fog();
        let clear_color = fog.clear_color;

        let mut uniform = ViewProjUniform::from_dmat4(&vp);
        uniform.fog_color = fog.fog_color;
        uniform.eye_pos = [eye.x as f32, eye.y as f32, eye.z as f32, 0.0];
        uniform.fog_params = [fog.fog_start, fog.fog_end, fog.fog_density, 0.0];
        if let Some(hillshade) = params.state.hillshade() {
            uniform.hillshade_highlight = hillshade.highlight_color;
            uniform.hillshade_shadow = hillshade.shadow_color;
            uniform.hillshade_accent = hillshade.accent_color;
            uniform.hillshade_light = [
                hillshade.illumination_direction,
                hillshade.illumination_altitude,
                hillshade.exaggeration,
                hillshade.opacity,
            ];
        }

        // Lighting: read pre-computed lighting from the engine.
        let lighting = params.state.computed_lighting();
        uniform.ambient_color = [
            lighting.ambient_color[0],
            lighting.ambient_color[1],
            lighting.ambient_color[2],
            lighting.lighting_enabled,
        ];
        uniform.directional_dir = [
            lighting.directional_dir[0],
            lighting.directional_dir[1],
            lighting.directional_dir[2],
            0.0,
        ];
        uniform.directional_color = [
            lighting.directional_color[0],
            lighting.directional_color[1],
            lighting.directional_color[2],
            0.0,
        ];

        params
            .queue
            .write_buffer(&self.uniform_buffer, 0, bytemuck::bytes_of(&uniform));

        // Upload sky uniform (separate buffer, different layout).
        {
            let sky = params.state.computed_sky();
            let inv_vp = vp.inverse();
            let inv_vp_f32 = inv_vp.as_mat4();
            let sky_uniform = SkyUniform {
                inv_view_proj: inv_vp_f32.to_cols_array_2d(),
                sun_dir: [
                    sky.sun_direction[0],
                    sky.sun_direction[1],
                    sky.sun_direction[2],
                    sky.sun_intensity,
                ],
                rayleigh_tint: [
                    sky.rayleigh_color[0],
                    sky.rayleigh_color[1],
                    sky.rayleigh_color[2],
                    sky.sky_enabled,
                ],
                mie_tint: [sky.mie_color[0], sky.mie_color[1], sky.mie_color[2], 1.0],
            };
            params.queue.write_buffer(
                &self.sky_uniform_buffer,
                0,
                bytemuck::bytes_of(&sky_uniform),
            );
        }

        // Upload shadow params uniform.
        {
            let shadow = params.state.computed_shadow();
            let shadow_uniform = ShadowParamsUniform {
                light_matrix_0: shadow.light_matrices[0],
                light_matrix_1: shadow.light_matrices[1],
                shadow_config: [
                    shadow.intensity,
                    shadow.texel_size,
                    shadow.normal_offset,
                    shadow.cascade_split,
                ],
                shadow_dir: [
                    lighting.directional_dir[0],
                    lighting.directional_dir[1],
                    lighting.directional_dir[2],
                    if shadow.enabled { 1.0 } else { 0.0 },
                ],
            };
            params.queue.write_buffer(
                &self.shadow_params_buffer,
                0,
                bytemuck::bytes_of(&shadow_uniform),
            );
        }

        // ?? 2. Enqueue new tile textures into the atlas ??????????????????
        for vt in params.visible_tiles {
            if let Some(TileData::Raster(ref img)) = vt.data {
                self.upload_tile(params.device, vt.actual, img);
            }
        }
        for raster in params.state.hillshade_rasters() {
            self.upload_hillshade(params.device, raster.tile, &raster.image);
        }

        // ?? 2b. Flush deferred atlas uploads (partial texture writes) ??
        self.flush_atlas_uploads(params.queue);

        // ?? 3. Mark visible + terrain tiles as used (prevents eviction) ?
        for vt in params.visible_tiles {
            self.tile_atlas.mark_used(&vt.actual);
        }
        let terrain_meshes = params.state.terrain_meshes();
        for mesh in terrain_meshes {
            if let Some(actual_tile) = find_terrain_texture_actual(mesh.tile, params.visible_tiles)
            {
                self.tile_atlas.mark_used(&actual_tile);
            }
        }
        for raster in params.state.hillshade_rasters() {
            self.hillshade_atlas.mark_used(&raster.tile);
        }

        let use_shared_terrain = !terrain_meshes.is_empty()
            && matches!(
                params.state.camera().projection(),
                rustial_engine::CameraProjection::WebMercator
                    | rustial_engine::CameraProjection::Equirectangular
            )
            && terrain_meshes
                .iter()
                .all(|mesh| mesh.elevation_texture.is_some());

        let materialized_terrain_meshes: Vec<TerrainMeshData> = if use_shared_terrain {
            Vec::new()
        } else {
            terrain_meshes
                .iter()
                .map(|mesh| {
                    materialize_terrain_mesh(
                        mesh,
                        params.state.camera().projection(),
                        rustial_engine::skirt_height(
                            mesh.tile.zoom,
                            mesh.vertical_exaggeration as f64,
                        ),
                    )
                })
                .collect()
        };

        // ?? 4. Cache model mesh GPU buffers ?????????????????????????????
        if !params.model_instances.is_empty() {
            self.cache_model_meshes(params.device, params.model_instances);
            self.cache_model_transforms(
                params.device,
                params.model_instances,
                scene_camera_origin,
                params.state,
            );
        } else {
            self.cached_model_transforms = None;
        }

        // ?? 5. Build batched geometry (tiles, terrain, vectors) ??????
        let tile_batch_key = TileBatchCacheKey::new(
            params.visible_tiles,
            scene_camera_origin,
            params.state.camera().projection(),
        );
        if self.tile_batch_cache_key.as_ref() != Some(&tile_batch_key) {
            self.cached_tile_batches = build_tile_batches(
                params.device,
                params.visible_tiles,
                &self.tile_atlas,
                scene_camera_origin,
                params.state.camera().projection(),
            );
            self.tile_batch_cache_key = Some(tile_batch_key);
        }

        let terrain_batches = if !use_shared_terrain && !materialized_terrain_meshes.is_empty() {
            build_terrain_batches(
                params.device,
                &materialized_terrain_meshes,
                &self.tile_atlas,
                scene_camera_origin,
                params.visible_tiles,
            )
        } else {
            Vec::new()
        };

        let hillshade_batches = if !materialized_terrain_meshes.is_empty()
            && !params.state.hillshade_rasters().is_empty()
        {
            build_hillshade_batches(
                params.device,
                &materialized_terrain_meshes,
                params.state.hillshade_rasters(),
                &self.hillshade_atlas,
                scene_camera_origin,
            )
        } else {
            Vec::new()
        };

        let vector_batch_key = VectorBatchCacheKey::new(params.vector_meshes, scene_camera_origin);
        if self.vector_batch_cache_key.as_ref() != Some(&vector_batch_key) {
            self.cached_vector_batches = params
                .vector_meshes
                .iter()
                .filter(|mesh| mesh.render_mode == VectorRenderMode::Generic)
                .map(|mesh| build_vector_batch(params.device, mesh, scene_camera_origin))
                .collect();
            self.cached_fill_batches = params
                .vector_meshes
                .iter()
                .filter(|mesh| {
                    mesh.render_mode == VectorRenderMode::Fill && mesh.fill_pattern.is_none()
                })
                .map(|mesh| {
                    build_fill_batch(
                        params.device,
                        mesh,
                        scene_camera_origin,
                        &self.uniform_buffer,
                        &self.fill_pipeline.uniform_bind_group_layout,
                    )
                })
                .collect();
            self.cached_fill_pattern_batches = params
                .vector_meshes
                .iter()
                .filter(|mesh| {
                    mesh.render_mode == VectorRenderMode::Fill && mesh.fill_pattern.is_some()
                })
                .map(|mesh| {
                    build_fill_pattern_batch(
                        params.device,
                        params.queue,
                        mesh,
                        scene_camera_origin,
                        &self.uniform_buffer,
                        &self.fill_pattern_pipeline.uniform_bind_group_layout,
                        &self.fill_pattern_pipeline.texture_bind_group_layout,
                        &self.fill_pattern_sampler,
                    )
                })
                .collect();
            self.cached_fill_extrusion_batches = params
                .vector_meshes
                .iter()
                .filter(|mesh| mesh.render_mode == VectorRenderMode::FillExtrusion)
                .map(|mesh| build_fill_extrusion_batch(params.device, mesh, scene_camera_origin))
                .collect();
            self.cached_line_batches = params
                .vector_meshes
                .iter()
                .filter(|mesh| {
                    mesh.render_mode == VectorRenderMode::Line && mesh.line_pattern.is_none()
                })
                .map(|mesh| build_line_batch(params.device, mesh, scene_camera_origin))
                .collect();
            self.cached_line_pattern_batches = params
                .vector_meshes
                .iter()
                .filter(|mesh| {
                    mesh.render_mode == VectorRenderMode::Line && mesh.line_pattern.is_some()
                })
                .map(|mesh| {
                    build_line_pattern_batch(
                        params.device,
                        params.queue,
                        mesh,
                        scene_camera_origin,
                        &self.uniform_buffer,
                        &self.line_pattern_pipeline.uniform_bind_group_layout,
                        &self.line_pattern_pipeline.texture_bind_group_layout,
                        &self.fill_pattern_sampler,
                    )
                })
                .collect();
            self.cached_circle_batches = params
                .vector_meshes
                .iter()
                .filter(|mesh| mesh.render_mode == VectorRenderMode::Circle)
                .map(|mesh| build_circle_batch(params.device, mesh, scene_camera_origin))
                .collect();
            self.cached_heatmap_batches = params
                .vector_meshes
                .iter()
                .filter(|mesh| mesh.render_mode == VectorRenderMode::Heatmap)
                .map(|mesh| build_heatmap_batch(params.device, mesh, scene_camera_origin))
                .collect();
            self.vector_batch_cache_key = Some(vector_batch_key);
        }

        // Build symbol batch from placed symbols.
        {
            let symbols = &frame.symbols;
            if !symbols.is_empty() {
                // Request glyphs for all visible symbols.
                self.symbol_glyph_atlas = rustial_engine::symbols::GlyphAtlas::new();
                for symbol in symbols.iter() {
                    if symbol.visible && symbol.opacity > 0.0 {
                        if let Some(text) = &symbol.text {
                            self.symbol_glyph_atlas
                                .request_text(&symbol.font_stack, text);
                        }
                    }
                }
                // Rasterize requested glyphs.
                self.symbol_glyph_atlas
                    .load_requested(&*self.symbol_glyph_provider);

                let dims = self.symbol_glyph_atlas.dimensions();
                if dims[0] > 0 && dims[1] > 0 {
                    // Upload atlas texture.
                    let tex = params.device.create_texture(&wgpu::TextureDescriptor {
                        label: Some("symbol_atlas_tex"),
                        size: wgpu::Extent3d {
                            width: dims[0] as u32,
                            height: dims[1] as u32,
                            depth_or_array_layers: 1,
                        },
                        mip_level_count: 1,
                        sample_count: 1,
                        dimension: wgpu::TextureDimension::D2,
                        format: wgpu::TextureFormat::R8Unorm,
                        usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::COPY_DST,
                        view_formats: &[],
                    });
                    params.queue.write_texture(
                        wgpu::TexelCopyTextureInfo {
                            texture: &tex,
                            mip_level: 0,
                            origin: wgpu::Origin3d::ZERO,
                            aspect: wgpu::TextureAspect::All,
                        },
                        self.symbol_glyph_atlas.alpha(),
                        wgpu::TexelCopyBufferLayout {
                            offset: 0,
                            bytes_per_row: Some(dims[0] as u32),
                            rows_per_image: Some(dims[1] as u32),
                        },
                        wgpu::Extent3d {
                            width: dims[0] as u32,
                            height: dims[1] as u32,
                            depth_or_array_layers: 1,
                        },
                    );
                    let view = tex.create_view(&wgpu::TextureViewDescriptor::default());
                    let atlas_bg = params.device.create_bind_group(&wgpu::BindGroupDescriptor {
                        label: Some("symbol_atlas_bg"),
                        layout: &self.symbol_pipeline.atlas_bind_group_layout,
                        entries: &[
                            wgpu::BindGroupEntry {
                                binding: 0,
                                resource: wgpu::BindingResource::TextureView(&view),
                            },
                            wgpu::BindGroupEntry {
                                binding: 1,
                                resource: wgpu::BindingResource::Sampler(&self.sampler),
                            },
                        ],
                    });
                    self.symbol_atlas_texture = Some((tex, view));
                    self.symbol_atlas_bind_group = Some(atlas_bg);
                }

                // Lay out glyph positions using atlas metrics.
                let mut laid_out_symbols: Vec<rustial_engine::symbols::PlacedSymbol> =
                    symbols.to_vec();
                rustial_engine::symbols::layout_symbol_glyphs(
                    &mut laid_out_symbols,
                    &self.symbol_glyph_atlas,
                );

                // Build the symbol geometry batch.
                self.cached_symbol_batch = build_symbol_batch(
                    params.device,
                    &laid_out_symbols,
                    &self.symbol_glyph_atlas,
                    scene_camera_origin,
                    self.symbol_glyph_atlas.render_em_px(),
                );
            } else {
                self.cached_symbol_batch = None;
                self.symbol_atlas_bind_group = None;
                self.symbol_atlas_texture = None;
            }
        }

        // Build placeholder quad batch for loading tiles.
        self.cached_placeholder_batch = build_placeholder_batches(
            params.device,
            &frame.placeholders,
            params.state.placeholder_style(),
            scene_camera_origin,
        );

        // Build image overlay batches.
        self.build_image_overlay_batches(
            params.device,
            params.queue,
            &frame.image_overlays,
            scene_camera_origin,
        );

        // Ensure shared terrain resources are prepared before render pass.
        if use_shared_terrain {
            for mesh in terrain_meshes {
                self.get_or_create_shared_grid(params.device, mesh.grid_resolution);
                let scene_origin = scene_camera_origin;
                // Prepare terrain tile bind groups for all three pipeline kinds.
                self.get_or_create_terrain_tile_bind(
                    params.device,
                    params.queue,
                    mesh,
                    params.state,
                    scene_origin,
                    TerrainPipelineKind::Terrain,
                );
                self.get_or_create_terrain_tile_bind(
                    params.device,
                    params.queue,
                    mesh,
                    params.state,
                    scene_origin,
                    TerrainPipelineKind::TerrainData,
                );
                self.get_or_create_terrain_tile_bind(
                    params.device,
                    params.queue,
                    mesh,
                    params.state,
                    scene_origin,
                    TerrainPipelineKind::Hillshade,
                );
            }
        }

        let grid_scalar_overlays: Vec<_> = visualization
            .iter()
            .filter(|overlay| matches!(overlay, VisualizationOverlay::GridScalar { .. }))
            .collect();
        let visible_grid_scalar_overlays: Vec<_> = grid_scalar_overlays
            .iter()
            .copied()
            .filter(|overlay| {
                visualization_overlay_intersects_scene_viewport(overlay, params.state)
            })
            .collect();
        let grid_extrusion_overlays: Vec<_> = visualization
            .iter()
            .filter(|overlay| matches!(overlay, VisualizationOverlay::GridExtrusion { .. }))
            .collect();
        let visible_grid_extrusion_overlays: Vec<_> = grid_extrusion_overlays
            .iter()
            .copied()
            .filter(|overlay| {
                visualization_overlay_intersects_scene_viewport(overlay, params.state)
            })
            .collect();
        let column_overlays: Vec<_> = visualization
            .iter()
            .filter(|overlay| matches!(overlay, VisualizationOverlay::Columns { .. }))
            .collect();
        let visible_column_overlays: Vec<_> = column_overlays
            .iter()
            .copied()
            .filter(|overlay| {
                visualization_overlay_intersects_scene_viewport(overlay, params.state)
            })
            .collect();
        let point_cloud_overlays: Vec<_> = visualization
            .iter()
            .filter(|overlay| matches!(overlay, VisualizationOverlay::Points { .. }))
            .collect();
        let visible_point_cloud_overlays: Vec<_> = point_cloud_overlays
            .iter()
            .copied()
            .filter(|overlay| {
                visualization_overlay_intersects_scene_viewport(overlay, params.state)
            })
            .collect();
        let terrain_fingerprint = TerrainDataDirtyState::terrain_fingerprint(terrain_meshes);
        if !visible_grid_scalar_overlays.is_empty() {
            for overlay in &visible_grid_scalar_overlays {
                self.get_or_create_grid_scalar_overlay(
                    params.device,
                    params.queue,
                    overlay,
                    params.state,
                    scene_camera_origin,
                    terrain_fingerprint,
                );
            }
        }
        if !visible_column_overlays.is_empty() {
            self.get_or_create_shared_column_mesh(params.device);
            for overlay in &visible_column_overlays {
                self.get_or_create_column_overlay(
                    params.device,
                    params.queue,
                    overlay,
                    params.state,
                    scene_camera_origin,
                );
            }
        }
        if !visible_point_cloud_overlays.is_empty() {
            self.get_or_create_shared_column_mesh(params.device);
            for overlay in &visible_point_cloud_overlays {
                self.get_or_create_point_cloud_overlay(
                    params.device,
                    params.queue,
                    overlay,
                    params.state,
                    scene_camera_origin,
                );
            }
        }
        if !visible_grid_extrusion_overlays.is_empty() {
            for overlay in &visible_grid_extrusion_overlays {
                self.get_or_create_grid_extrusion_overlay(
                    params.device,
                    params.queue,
                    overlay,
                    params.state,
                    scene_camera_origin,
                    terrain_fingerprint,
                );
            }
        }

        // ?? 6. Render pass ??????????????????????????????????????????????
        let mut encoder = params
            .device
            .create_command_encoder(&wgpu::CommandEncoderDescriptor {
                label: Some("rustial_encoder"),
            });

        let has_heatmap = !self.cached_heatmap_batches.is_empty()
            && self.cached_heatmap_batches.iter().any(|b| b.is_some());

        let painter_plan = PainterPlan::with_shadows(
            !terrain_meshes.is_empty(),
            !params.state.hillshade_rasters().is_empty(),
            has_heatmap,
            params.state.computed_shadow().enabled,
        );

        for painter_pass in painter_plan.iter() {
            match painter_pass {
                PainterPass::SkyAtmosphere => {
                    let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
                        label: Some("rustial_pass_sky_atmosphere"),
                        color_attachments: &[Some(wgpu::RenderPassColorAttachment {
                            view: params.color_view,
                            resolve_target: None,
                            ops: wgpu::Operations {
                                load: wgpu::LoadOp::Clear(wgpu::Color {
                                    r: clear_color[0] as f64,
                                    g: clear_color[1] as f64,
                                    b: clear_color[2] as f64,
                                    a: clear_color[3] as f64,
                                }),
                                store: wgpu::StoreOp::Store,
                            },
                        })],
                        depth_stencil_attachment: Some(wgpu::RenderPassDepthStencilAttachment {
                            view: &self.depth_view,
                            depth_ops: Some(wgpu::Operations {
                                load: wgpu::LoadOp::Clear(1.0),
                                store: wgpu::StoreOp::Store,
                            }),
                            stencil_ops: None,
                        }),
                        ..Default::default()
                    });

                    // Draw the atmosphere sky (fullscreen triangle at depth 1.0).
                    // The sky shader discards when sky_enabled < 0.5.
                    if params.state.computed_sky().sky_enabled > 0.5 {
                        pass.set_pipeline(&self.sky_pipeline.pipeline);
                        pass.set_bind_group(0, &self.sky_uniform_bind_group, &[]);
                        pass.draw(0..3, 0..1);
                    }
                }
                PainterPass::TerrainData => {
                    // Conditional terrain-data refresh: skip redrawing
                    // the depth / coordinate interaction buffers when the
                    // view-projection matrix and the terrain tile set
                    // haven't changed since the last render.  This mirrors
                    // MapLibre's `maybeDrawDepthAndCoords(false)` pattern.
                    if !self.terrain_data_dirty.needs_update(&vp, terrain_meshes) {
                        continue;
                    }
                    {
                        let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
                            label: Some("rustial_pass_terrain_data"),
                            color_attachments: &[Some(wgpu::RenderPassColorAttachment {
                                view: self.terrain_interaction_buffers.coord_view(),
                                resolve_target: None,
                                ops: wgpu::Operations {
                                    load: wgpu::LoadOp::Clear(wgpu::Color::TRANSPARENT),
                                    store: wgpu::StoreOp::Store,
                                },
                            })],
                            depth_stencil_attachment: Some(
                                wgpu::RenderPassDepthStencilAttachment {
                                    view: self.terrain_interaction_buffers.depth_view(),
                                    depth_ops: Some(wgpu::Operations {
                                        load: wgpu::LoadOp::Clear(1.0),
                                        store: wgpu::StoreOp::Store,
                                    }),
                                    stencil_ops: None,
                                },
                            ),
                            ..Default::default()
                        });
                        if use_shared_terrain {
                            self.render_shared_terrain_data_tiles(
                                &mut pass,
                                params.state,
                                terrain_meshes,
                            );
                        } else {
                            self.render_terrain_data_batches(&mut pass, &terrain_batches);
                        }
                    }
                    self.terrain_data_dirty.mark_clean(&vp, terrain_meshes);
                }
                PainterPass::ShadowDepth => {
                    let shadow = params.state.computed_shadow();
                    let cascade_count = (shadow.cascade_count as usize)
                        .min(self.shadow_map_views.len())
                        .min(self.shadow_depth_bind_groups.len());

                    // Upload all cascade light VPs upfront.
                    for cascade_idx in 0..cascade_count {
                        let light_vp_bytes: &[u8] =
                            bytemuck::cast_slice(&shadow.light_matrices[cascade_idx]);
                        params.queue.write_buffer(
                            &self.shadow_depth_uniform_buffers[cascade_idx],
                            0,
                            light_vp_bytes,
                        );
                    }

                    for cascade_idx in 0..cascade_count {
                        // Shadow depth render pass (depth-only, no colour attachment).
                        let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
                            label: Some("rustial_pass_shadow_depth"),
                            color_attachments: &[],
                            depth_stencil_attachment: Some(
                                wgpu::RenderPassDepthStencilAttachment {
                                    view: &self.shadow_map_views[cascade_idx],
                                    depth_ops: Some(wgpu::Operations {
                                        load: wgpu::LoadOp::Clear(1.0),
                                        store: wgpu::StoreOp::Store,
                                    }),
                                    stencil_ops: None,
                                },
                            ),
                            ..Default::default()
                        });

                        // Render fill-extrusion casters into shadow map.
                        if self
                            .cached_fill_extrusion_batches
                            .iter()
                            .any(|b| b.is_some())
                        {
                            pass.set_pipeline(&self.shadow_fill_extrusion_pipeline.pipeline);
                            pass.set_bind_group(
                                0,
                                &self.shadow_depth_bind_groups[cascade_idx],
                                &[],
                            );
                            for batch in self.cached_fill_extrusion_batches.iter().flatten() {
                                pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
                                pass.set_index_buffer(
                                    batch.index_buffer.slice(..),
                                    wgpu::IndexFormat::Uint32,
                                );
                                pass.draw_indexed(0..batch.index_count, 0, 0..1);
                            }
                        }

                        // Render model casters into shadow map.
                        if let Some(ref cached) = self.cached_model_transforms {
                            if cached.instance_count > 0 {
                                pass.set_pipeline(&self.shadow_model_pipeline.pipeline);
                                pass.set_bind_group(
                                    0,
                                    &self.shadow_depth_bind_groups[cascade_idx],
                                    &[],
                                );
                                for i in 0..cached.instance_count {
                                    let dyn_offset = (i * cached.stride) as u32;
                                    if let Some(instance) = params.model_instances.get(i) {
                                        let mesh_key = ModelMeshKey::from_mesh(&instance.mesh);
                                        if let Some(cached_mesh) =
                                            self.model_mesh_cache.get(&mesh_key)
                                        {
                                            pass.set_bind_group(
                                                1,
                                                &cached.bind_group,
                                                &[dyn_offset],
                                            );
                                            pass.set_vertex_buffer(
                                                0,
                                                cached_mesh.vertex_buffer.slice(..),
                                            );
                                            pass.set_index_buffer(
                                                cached_mesh.index_buffer.slice(..),
                                                wgpu::IndexFormat::Uint32,
                                            );
                                            pass.draw_indexed(0..cached_mesh.index_count, 0, 0..1);
                                        }
                                    }
                                }
                            }
                        }
                    }
                }
                PainterPass::OpaqueScene => {
                    let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
                        label: Some("rustial_pass_opaque"),
                        color_attachments: &[Some(wgpu::RenderPassColorAttachment {
                            view: params.color_view,
                            resolve_target: None,
                            ops: wgpu::Operations {
                                load: wgpu::LoadOp::Load,
                                store: wgpu::StoreOp::Store,
                            },
                        })],
                        depth_stencil_attachment: Some(wgpu::RenderPassDepthStencilAttachment {
                            view: &self.depth_view,
                            depth_ops: Some(wgpu::Operations {
                                load: wgpu::LoadOp::Load,
                                store: wgpu::StoreOp::Store,
                            }),
                            stencil_ops: None,
                        }),
                        ..Default::default()
                    });

                    // Draw loading placeholder quads behind tiles so
                    // loaded imagery naturally occludes them.
                    if let Some(ref ph_batch) = self.cached_placeholder_batch {
                        pass.set_pipeline(&self.vector_pipeline.pipeline);
                        pass.set_bind_group(0, &self.vector_uniform_bind_group, &[]);
                        pass.set_vertex_buffer(0, ph_batch.vertex_buffer.slice(..));
                        pass.set_index_buffer(
                            ph_batch.index_buffer.slice(..),
                            wgpu::IndexFormat::Uint32,
                        );
                        pass.draw_indexed(0..ph_batch.index_count, 0, 0..1);
                    }

                    if use_shared_terrain {
                        self.render_shared_terrain_tiles(
                            &mut pass,
                            params.state,
                            terrain_meshes,
                            params.visible_tiles,
                        );
                    } else if !terrain_batches.is_empty() {
                        self.render_terrain_batches(&mut pass, &terrain_batches);
                    } else {
                        self.render_tile_batches(&mut pass, &self.cached_tile_batches);
                    }

                    if !visible_grid_scalar_overlays.is_empty() {
                        self.render_grid_scalar_overlays(&mut pass, &visible_grid_scalar_overlays);
                    }
                    if !visible_grid_extrusion_overlays.is_empty() {
                        self.render_grid_extrusion_overlays(
                            &mut pass,
                            &visible_grid_extrusion_overlays,
                        );
                    }
                    if !visible_column_overlays.is_empty() {
                        self.render_column_overlays(&mut pass, &visible_column_overlays);
                    }
                    if !visible_point_cloud_overlays.is_empty() {
                        self.render_point_cloud_overlays(&mut pass, &visible_point_cloud_overlays);
                    }

                    self.render_vector_batches(&mut pass, &self.cached_vector_batches);
                    self.render_fill_batches(&mut pass, &self.cached_fill_batches);
                    self.render_fill_pattern_batches(&mut pass, &self.cached_fill_pattern_batches);
                    self.render_fill_extrusion_batches(
                        &mut pass,
                        &self.cached_fill_extrusion_batches,
                    );
                    self.render_line_batches(&mut pass, &self.cached_line_batches);
                    self.render_line_pattern_batches(&mut pass, &self.cached_line_pattern_batches);
                    self.render_circle_batches(&mut pass, &self.cached_circle_batches);
                    // Heatmap is now rendered in its own two-pass pipeline
                    // (HeatmapAccumulation → HeatmapColormap) rather than
                    // directly into the opaque scene.
                    self.render_image_overlay_batches(&mut pass);
                    self.render_symbol_batch(&mut pass);

                    if !params.model_instances.is_empty() {
                        self.render_models(&mut pass, params.model_instances, params.device);
                    }
                }
                PainterPass::HeatmapAccumulation => {
                    // Pass 1: Render Gaussian-weighted heatmap points into
                    // the off-screen R16Float accumulation texture.
                    let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
                        label: Some("rustial_pass_heatmap_accum"),
                        color_attachments: &[Some(wgpu::RenderPassColorAttachment {
                            view: &self.heatmap_accum_view,
                            resolve_target: None,
                            ops: wgpu::Operations {
                                load: wgpu::LoadOp::Clear(wgpu::Color {
                                    r: 0.0,
                                    g: 0.0,
                                    b: 0.0,
                                    a: 0.0,
                                }),
                                store: wgpu::StoreOp::Store,
                            },
                        })],
                        depth_stencil_attachment: Some(wgpu::RenderPassDepthStencilAttachment {
                            view: &self.depth_view,
                            depth_ops: Some(wgpu::Operations {
                                load: wgpu::LoadOp::Load,
                                store: wgpu::StoreOp::Store,
                            }),
                            stencil_ops: None,
                        }),
                        ..Default::default()
                    });
                    self.render_heatmap_batches(&mut pass, &self.cached_heatmap_batches);
                }
                PainterPass::HeatmapColormap => {
                    // Pass 2: Fullscreen triangle reads the accumulated
                    // weight texture and maps it through a colour ramp,
                    // compositing onto the main surface.
                    let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
                        label: Some("rustial_pass_heatmap_colormap"),
                        color_attachments: &[Some(wgpu::RenderPassColorAttachment {
                            view: params.color_view,
                            resolve_target: None,
                            ops: wgpu::Operations {
                                load: wgpu::LoadOp::Load,
                                store: wgpu::StoreOp::Store,
                            },
                        })],
                        depth_stencil_attachment: None,
                        ..Default::default()
                    });
                    pass.set_pipeline(&self.heatmap_colormap_pipeline.pipeline);
                    pass.set_bind_group(0, &self.heatmap_colormap_uniform_bind_group, &[]);
                    pass.set_bind_group(1, &self.heatmap_colormap_textures_bind_group, &[]);
                    pass.draw(0..3, 0..1); // fullscreen triangle
                }
                PainterPass::HillshadeOverlay => {
                    let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
                        label: Some("rustial_pass_hillshade_overlay"),
                        color_attachments: &[Some(wgpu::RenderPassColorAttachment {
                            view: params.color_view,
                            resolve_target: None,
                            ops: wgpu::Operations {
                                load: wgpu::LoadOp::Load,
                                store: wgpu::StoreOp::Store,
                            },
                        })],
                        depth_stencil_attachment: Some(wgpu::RenderPassDepthStencilAttachment {
                            view: &self.depth_view,
                            depth_ops: Some(wgpu::Operations {
                                load: wgpu::LoadOp::Load,
                                store: wgpu::StoreOp::Store,
                            }),
                            stencil_ops: None,
                        }),
                        ..Default::default()
                    });

                    if use_shared_terrain {
                        self.render_shared_hillshade_tiles(&mut pass, params.state, terrain_meshes);
                    } else {
                        self.render_hillshade_batches(&mut pass, &hillshade_batches);
                    }
                }
            }
        }

        // ?? 7. Submit ???????????????????????????????????????????????????
        params.queue.submit(std::iter::once(encoder.finish()));

        self.prune_height_texture_cache(terrain_meshes);
        self.prune_terrain_tile_bind_cache(terrain_meshes);
        self.prune_grid_scalar_overlay_cache(&grid_scalar_overlays);
        self.prune_grid_extrusion_overlay_cache(&grid_extrusion_overlays);
        self.prune_column_overlay_cache(&column_overlays);
        self.prune_point_cloud_overlay_cache(&point_cloud_overlays);

        // ?? 8. Atlas end-of-frame eviction ??????????????????????????????
        let tile_count_before = self.tile_atlas.len();
        self.tile_atlas.end_frame();
        if self.tile_atlas.len() != tile_count_before {
            self.tile_batch_cache_key = None;
        }
        self.hillshade_atlas.end_frame();
    }

    // -- Batched tile rendering -------------------------------------------

    /// Issue one `draw_indexed` per atlas page that has visible tile geometry.
    fn render_tile_batches<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        batches: &'a [TilePageBatches],
    ) {
        pass.set_bind_group(0, &self.uniform_bind_group, &[]);

        for (page_idx, batch) in batches.iter().enumerate() {
            let bg = match self.page_bind_groups.get(page_idx) {
                Some(bg) => bg,
                None => continue,
            };

            if let Some(batch) = batch.opaque.as_ref() {
                pass.set_pipeline(&self.tile_pipeline.pipeline);
                pass.set_bind_group(1, bg, &[]);
                pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
                pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
                pass.draw_indexed(0..batch.index_count, 0, 0..1);
            }

            if let Some(batch) = batch.translucent.as_ref() {
                pass.set_pipeline(&self.tile_pipeline.translucent_pipeline);
                pass.set_bind_group(1, bg, &[]);
                pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
                pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
                pass.draw_indexed(0..batch.index_count, 0, 0..1);
            }
        }
    }

    fn render_grid_scalar_overlays<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        overlays: &[&'a VisualizationOverlay],
    ) {
        pass.set_pipeline(&self.grid_scalar_pipeline.pipeline);
        pass.set_bind_group(0, &self.grid_scalar_uniform_bind_group, &[]);

        for overlay in overlays {
            let VisualizationOverlay::GridScalar { layer_id, .. } = overlay else {
                continue;
            };
            let Some(cached) = self.grid_scalar_overlay_cache.get(layer_id) else {
                continue;
            };
            pass.set_bind_group(1, &cached.bind_group, &[]);
            pass.set_vertex_buffer(0, cached.vertex_buffer.slice(..));
            pass.set_index_buffer(cached.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..cached.index_count, 0, 0..1);
        }
    }

    fn render_grid_extrusion_overlays<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        overlays: &[&'a VisualizationOverlay],
    ) {
        pass.set_pipeline(&self.grid_extrusion_pipeline.pipeline);
        pass.set_bind_group(0, &self.grid_extrusion_uniform_bind_group, &[]);

        for overlay in overlays {
            let VisualizationOverlay::GridExtrusion { layer_id, .. } = overlay else {
                continue;
            };
            let Some(cached) = self.grid_extrusion_overlay_cache.get(layer_id) else {
                continue;
            };
            pass.set_vertex_buffer(0, cached.vertex_buffer.slice(..));
            pass.set_index_buffer(cached.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..cached.index_count, 0, 0..1);
        }
    }

    fn render_column_overlays<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        overlays: &[&'a VisualizationOverlay],
    ) {
        let Some(mesh) = self.shared_column_mesh.as_ref() else {
            return;
        };

        pass.set_pipeline(&self.column_pipeline.pipeline);
        pass.set_bind_group(0, &self.column_uniform_bind_group, &[]);

        for overlay in overlays {
            let VisualizationOverlay::Columns { layer_id, .. } = overlay else {
                continue;
            };
            let Some(cached) = self.column_overlay_cache.get(layer_id) else {
                continue;
            };
            if cached.instance_count == 0 {
                continue;
            }
            pass.set_vertex_buffer(0, mesh.vertex_buffer.slice(..));
            pass.set_vertex_buffer(1, cached.instance_buffer.slice(..));
            pass.set_index_buffer(mesh.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..mesh.index_count, 0, 0..cached.instance_count);
        }
    }

    fn render_point_cloud_overlays<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        overlays: &[&'a VisualizationOverlay],
    ) {
        let Some(mesh) = self.shared_column_mesh.as_ref() else {
            return;
        };

        pass.set_pipeline(&self.column_pipeline.pipeline);
        pass.set_bind_group(0, &self.column_uniform_bind_group, &[]);

        for overlay in overlays {
            let VisualizationOverlay::Points { layer_id, .. } = overlay else {
                continue;
            };
            let Some(cached) = self.point_cloud_overlay_cache.get(layer_id) else {
                continue;
            };
            if cached.instance_count == 0 {
                continue;
            }
            pass.set_vertex_buffer(0, mesh.vertex_buffer.slice(..));
            pass.set_vertex_buffer(1, cached.instance_buffer.slice(..));
            pass.set_index_buffer(mesh.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..mesh.index_count, 0, 0..cached.instance_count);
        }
    }

    // -- Batched terrain rendering ----------------------------------------

    /// Issue one `draw_indexed` per atlas page that has terrain geometry.
    fn render_terrain_batches<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        batches: &'a [Option<TerrainBatch>],
    ) {
        pass.set_pipeline(&self.terrain_pipeline.pipeline);
        pass.set_bind_group(0, &self.terrain_uniform_bind_group, &[]);

        for (page_idx, batch) in batches.iter().enumerate() {
            let batch = match batch {
                Some(b) => b,
                None => continue,
            };
            let bg = match self.page_terrain_bind_groups.get(page_idx) {
                Some(bg) => bg,
                None => continue,
            };

            pass.set_bind_group(1, bg, &[]);
            pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
            pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..batch.index_count, 0, 0..1);
        }
    }

    /// Issue one `draw_indexed` per terrain batch into the renderer-owned
    /// depth / coordinate buffers.
    fn render_terrain_data_batches<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        batches: &'a [Option<TerrainBatch>],
    ) {
        pass.set_pipeline(&self.terrain_data_pipeline.pipeline);
        pass.set_bind_group(0, &self.terrain_data_uniform_bind_group, &[]);

        for batch in batches.iter().flatten() {
            pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
            pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..batch.index_count, 0, 0..1);
        }
    }

    // -- Batched hillshade rendering -------------------------------------

    /// Issue one `draw_indexed` per atlas page that has hillshade geometry.
    fn render_hillshade_batches<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        batches: &'a [Option<HillshadeBatch>],
    ) {
        pass.set_pipeline(&self.hillshade_pipeline.pipeline);
        pass.set_bind_group(0, &self.hillshade_uniform_bind_group, &[]);

        for (page_idx, batch) in batches.iter().enumerate() {
            let batch = match batch {
                Some(b) => b,
                None => continue,
            };
            let bg = match self.page_hillshade_bind_groups.get(page_idx) {
                Some(bg) => bg,
                None => continue,
            };

            pass.set_bind_group(1, bg, &[]);
            pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
            pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..batch.index_count, 0, 0..1);
        }
    }

    // -- Batched vector rendering -----------------------------------------

    /// Draw each non-empty vector layer.  Pipeline state is set lazily on
    /// the first actual draw to avoid overhead when there are no vectors.
    fn render_vector_batches<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        batches: &'a [Option<VectorBatchEntry>],
    ) {
        let mut pipeline_set = false;

        for batch in batches.iter().flatten() {
            if !pipeline_set {
                pass.set_pipeline(&self.vector_pipeline.pipeline);
                pass.set_bind_group(0, &self.vector_uniform_bind_group, &[]);
                pipeline_set = true;
            }

            pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
            pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..batch.index_count, 0, 0..1);
        }
    }

    // -- Batched fill rendering -------------------------------------------

    /// Draw each non-empty fill layer with per-batch fill params.
    fn render_fill_batches<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        batches: &'a [Option<FillBatchEntry>],
    ) {
        for batch in batches.iter().flatten() {
            pass.set_pipeline(&self.fill_pipeline.pipeline);
            // Each batch has its own bind group (view-proj + fill params).
            pass.set_bind_group(0, &batch.bind_group, &[]);
            pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
            pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..batch.index_count, 0, 0..1);
        }
    }

    // -- Batched fill-pattern rendering ------------------------------------

    /// Draw each non-empty fill-pattern layer with texture sampling.
    fn render_fill_pattern_batches<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        batches: &'a [Option<FillPatternBatchEntry>],
    ) {
        for batch in batches.iter().flatten() {
            pass.set_pipeline(&self.fill_pattern_pipeline.pipeline);
            pass.set_bind_group(0, &batch.uniform_bind_group, &[]);
            pass.set_bind_group(1, &batch.texture_bind_group, &[]);
            pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
            pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..batch.index_count, 0, 0..1);
        }
    }

    // -- Batched fill-extrusion rendering ---------------------------------

    /// Draw each non-empty fill-extrusion layer with per-face lighting.
    fn render_fill_extrusion_batches<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        batches: &'a [Option<FillExtrusionBatchEntry>],
    ) {
        let mut pipeline_set = false;

        for batch in batches.iter().flatten() {
            if !pipeline_set {
                pass.set_pipeline(&self.fill_extrusion_pipeline.pipeline);
                pass.set_bind_group(0, &self.fill_extrusion_uniform_bind_group, &[]);
                pass.set_bind_group(1, &self.shadow_receiver_bind_group, &[]);
                pipeline_set = true;
            }

            pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
            pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..batch.index_count, 0, 0..1);
        }
    }

    fn render_line_batches<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        batches: &'a [Option<LineBatchEntry>],
    ) {
        let mut pipeline_set = false;

        for batch in batches.iter().flatten() {
            if !pipeline_set {
                pass.set_pipeline(&self.line_pipeline.pipeline);
                pass.set_bind_group(0, &self.line_uniform_bind_group, &[]);
                pipeline_set = true;
            }

            pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
            pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..batch.index_count, 0, 0..1);
        }
    }

    fn render_line_pattern_batches<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        batches: &'a [Option<LinePatternBatchEntry>],
    ) {
        for batch in batches.iter().flatten() {
            pass.set_pipeline(&self.line_pattern_pipeline.pipeline);
            pass.set_bind_group(0, &batch.uniform_bind_group, &[]);
            pass.set_bind_group(1, &batch.texture_bind_group, &[]);
            pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
            pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..batch.index_count, 0, 0..1);
        }
    }

    fn render_circle_batches<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        batches: &'a [Option<CircleBatchEntry>],
    ) {
        let mut pipeline_set = false;

        for batch in batches.iter().flatten() {
            if !pipeline_set {
                pass.set_pipeline(&self.circle_pipeline.pipeline);
                pass.set_bind_group(0, &self.circle_uniform_bind_group, &[]);
                pipeline_set = true;
            }

            pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
            pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..batch.index_count, 0, 0..1);
        }
    }

    fn render_heatmap_batches<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        batches: &'a [Option<HeatmapBatchEntry>],
    ) {
        let mut pipeline_set = false;

        for batch in batches.iter().flatten() {
            if !pipeline_set {
                pass.set_pipeline(&self.heatmap_pipeline.pipeline);
                pass.set_bind_group(0, &self.heatmap_uniform_bind_group, &[]);
                pipeline_set = true;
            }

            pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
            pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..batch.index_count, 0, 0..1);
        }
    }

    // -- Image overlay batches ---------------------------------------------

    /// Build GPU batches for image overlays.  Reuses textures and bind
    /// groups from the previous frame when the overlay data has not
    /// changed (Arc pointer identity) or when only the geometry moved
    /// (same dimensions → `write_texture` instead of recreating).
    fn build_image_overlay_batches(
        &mut self,
        device: &wgpu::Device,
        queue: &wgpu::Queue,
        overlays: &[rustial_engine::layers::ImageOverlayData],
        camera_origin: glam::DVec3,
    ) {
        // Take the old cache so we can harvest reusable entries.
        let mut old_cache: Vec<CachedImageOverlayBatch> =
            std::mem::take(&mut self.cached_image_overlay_batches);

        for overlay in overlays {
            if overlay.width == 0 || overlay.height == 0 || overlay.data.is_empty() {
                continue;
            }

            let data_arc_ptr = Arc::as_ptr(&overlay.data) as usize;

            // Try to find a matching cached entry (by layer id).
            let cached_idx = old_cache
                .iter()
                .position(|c| c.layer_id == overlay.layer_id);

            // Corner UVs: [top-left, top-right, bottom-right, bottom-left]
            let uvs = [[0.0, 0.0], [1.0, 0.0], [1.0, 1.0], [0.0, 1.0]];
            let vertices: Vec<ImageOverlayVertex> = overlay
                .corners
                .iter()
                .zip(uvs.iter())
                .map(|(corner, uv)| {
                    let rel = [
                        (corner[0] - camera_origin.x) as f32,
                        (corner[1] - camera_origin.y) as f32,
                        (corner[2] - camera_origin.z) as f32,
                    ];
                    ImageOverlayVertex {
                        position: rel,
                        uv: *uv,
                        opacity: overlay.opacity,
                    }
                })
                .collect();
            let indices: Vec<u32> = vec![0, 1, 2, 0, 2, 3];

            if let Some(idx) = cached_idx {
                let mut cached = old_cache.remove(idx);

                // Always rebuild vertex/index (positions may have changed).
                cached.vertex_buffer =
                    device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                        label: Some("image_overlay_vb"),
                        contents: bytemuck::cast_slice(&vertices),
                        usage: wgpu::BufferUsages::VERTEX,
                    });
                cached.index_buffer =
                    device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                        label: Some("image_overlay_ib"),
                        contents: bytemuck::cast_slice(&indices),
                        usage: wgpu::BufferUsages::INDEX,
                    });

                // If data pointer is identical, skip texture re-upload.
                if cached.data_arc_ptr != data_arc_ptr {
                    if cached.tex_dimensions == (overlay.width, overlay.height) {
                        // Same dimensions → write_texture (no reallocation).
                        queue.write_texture(
                            wgpu::TexelCopyTextureInfo {
                                texture: &cached.texture,
                                mip_level: 0,
                                origin: wgpu::Origin3d::ZERO,
                                aspect: wgpu::TextureAspect::All,
                            },
                            &overlay.data,
                            wgpu::TexelCopyBufferLayout {
                                offset: 0,
                                bytes_per_row: Some(overlay.width * 4),
                                rows_per_image: Some(overlay.height),
                            },
                            wgpu::Extent3d {
                                width: overlay.width,
                                height: overlay.height,
                                depth_or_array_layers: 1,
                            },
                        );
                    } else {
                        // Dimensions changed → recreate texture + bind group.
                        let (texture, texture_view, texture_bind_group) =
                            self.create_overlay_texture(device, queue, overlay);
                        cached.texture = texture;
                        cached.texture_view = texture_view;
                        cached.texture_bind_group = texture_bind_group;
                        cached.tex_dimensions = (overlay.width, overlay.height);
                    }
                    cached.data_arc_ptr = data_arc_ptr;
                }

                self.cached_image_overlay_batches.push(cached);
            } else {
                // New overlay — full creation.
                let vertex_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                    label: Some("image_overlay_vb"),
                    contents: bytemuck::cast_slice(&vertices),
                    usage: wgpu::BufferUsages::VERTEX,
                });
                let index_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                    label: Some("image_overlay_ib"),
                    contents: bytemuck::cast_slice(&indices),
                    usage: wgpu::BufferUsages::INDEX,
                });
                let (texture, texture_view, texture_bind_group) =
                    self.create_overlay_texture(device, queue, overlay);

                self.cached_image_overlay_batches
                    .push(CachedImageOverlayBatch {
                        vertex_buffer,
                        index_buffer,
                        texture,
                        texture_view,
                        texture_bind_group,
                        layer_id: overlay.layer_id,
                        tex_dimensions: (overlay.width, overlay.height),
                        data_arc_ptr,
                    });
            }
        }
        // old_cache entries not matched are dropped (stale overlays).
    }

    /// Create a new GPU texture + bind group for an image overlay.
    fn create_overlay_texture(
        &self,
        device: &wgpu::Device,
        queue: &wgpu::Queue,
        overlay: &rustial_engine::layers::ImageOverlayData,
    ) -> (wgpu::Texture, wgpu::TextureView, wgpu::BindGroup) {
        let texture = device.create_texture(&wgpu::TextureDescriptor {
            label: Some("image_overlay_tex"),
            size: wgpu::Extent3d {
                width: overlay.width,
                height: overlay.height,
                depth_or_array_layers: 1,
            },
            mip_level_count: 1,
            sample_count: 1,
            dimension: wgpu::TextureDimension::D2,
            format: wgpu::TextureFormat::Rgba8UnormSrgb,
            usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::COPY_DST,
            view_formats: &[],
        });
        queue.write_texture(
            wgpu::TexelCopyTextureInfo {
                texture: &texture,
                mip_level: 0,
                origin: wgpu::Origin3d::ZERO,
                aspect: wgpu::TextureAspect::All,
            },
            &overlay.data,
            wgpu::TexelCopyBufferLayout {
                offset: 0,
                bytes_per_row: Some(overlay.width * 4),
                rows_per_image: Some(overlay.height),
            },
            wgpu::Extent3d {
                width: overlay.width,
                height: overlay.height,
                depth_or_array_layers: 1,
            },
        );
        let texture_view = texture.create_view(&wgpu::TextureViewDescriptor::default());
        let texture_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("image_overlay_tex_bg"),
            layout: &self.image_overlay_pipeline.texture_bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: wgpu::BindingResource::TextureView(&texture_view),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: wgpu::BindingResource::Sampler(&self.sampler),
                },
            ],
        });
        (texture, texture_view, texture_bind_group)
    }

    /// Draw all cached image overlay batches.
    fn render_image_overlay_batches<'a>(&'a self, pass: &mut wgpu::RenderPass<'a>) {
        if self.cached_image_overlay_batches.is_empty() {
            return;
        }
        pass.set_pipeline(&self.image_overlay_pipeline.pipeline);
        pass.set_bind_group(0, &self.image_overlay_uniform_bind_group, &[]);
        for batch in &self.cached_image_overlay_batches {
            pass.set_bind_group(1, &batch.texture_bind_group, &[]);
            pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
            pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
            pass.draw_indexed(0..6, 0, 0..1);
        }
    }

    fn render_symbol_batch<'a>(&'a self, pass: &mut wgpu::RenderPass<'a>) {
        let batch = match &self.cached_symbol_batch {
            Some(b) => b,
            None => return,
        };
        let atlas_bg = match &self.symbol_atlas_bind_group {
            Some(bg) => bg,
            None => return,
        };

        pass.set_pipeline(&self.symbol_pipeline.pipeline);
        pass.set_bind_group(0, &self.symbol_uniform_bind_group, &[]);
        pass.set_bind_group(1, atlas_bg, &[]);
        pass.set_vertex_buffer(0, batch.vertex_buffer.slice(..));
        pass.set_index_buffer(batch.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
        pass.draw_indexed(0..batch.index_count, 0, 0..1);
    }

    fn render_models<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        model_instances: &[ModelInstance],
        device: &wgpu::Device,
    ) {
        if model_instances.is_empty() {
            return;
        }

        let cached = match &self.cached_model_transforms {
            Some(c) if c.instance_count == model_instances.len() => c,
            _ => return,
        };

        pass.set_pipeline(&self.model_pipeline.pipeline);
        pass.set_bind_group(0, &self.model_uniform_bind_group, &[]);
        pass.set_bind_group(2, &self.shadow_receiver_bind_group, &[]);

        for (i, instance) in model_instances.iter().enumerate() {
            let dyn_offset = (i * cached.stride) as u32;
            let mesh_key = ModelMeshKey::from_mesh(&instance.mesh);
            let cached_mesh = self.model_mesh_cache.get(&mesh_key);

            if let Some(cached_mesh) = cached_mesh {
                pass.set_bind_group(1, &cached.bind_group, &[dyn_offset]);
                pass.set_vertex_buffer(0, cached_mesh.vertex_buffer.slice(..));
                pass.set_index_buffer(
                    cached_mesh.index_buffer.slice(..),
                    wgpu::IndexFormat::Uint32,
                );
                pass.draw_indexed(0..cached_mesh.index_count, 0, 0..1);
            } else {
                let vertices = build_model_vertices(&instance.mesh);
                let vb = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                    label: Some("model_inline_vb"),
                    contents: bytemuck::cast_slice(&vertices),
                    usage: wgpu::BufferUsages::VERTEX,
                });
                let ib = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                    label: Some("model_inline_ib"),
                    contents: bytemuck::cast_slice(&instance.mesh.indices),
                    usage: wgpu::BufferUsages::INDEX,
                });
                let index_count = instance.mesh.indices.len() as u32;
                pass.set_bind_group(1, &cached.bind_group, &[dyn_offset]);
                pass.set_vertex_buffer(0, vb.slice(..));
                pass.set_index_buffer(ib.slice(..), wgpu::IndexFormat::Uint32);
                pass.draw_indexed(0..index_count, 0, 0..1);
            }
        }
    }

    /// Pre-cache model mesh GPU buffers before the render pass begins.
    ///
    /// Called automatically by [`render_full`](Self::render_full). This keeps
    /// stable model meshes resident on the GPU instead of re-uploading them
    /// every frame.
    pub fn cache_model_meshes(&mut self, device: &wgpu::Device, model_instances: &[ModelInstance]) {
        for instance in model_instances {
            let key = ModelMeshKey::from_mesh(&instance.mesh);
            if self.model_mesh_cache.contains_key(&key) {
                continue;
            }

            let vertices = build_model_vertices(&instance.mesh);
            let vb = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("cached_model_vb"),
                contents: bytemuck::cast_slice(&vertices),
                usage: wgpu::BufferUsages::VERTEX,
            });
            let ib = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("cached_model_ib"),
                contents: bytemuck::cast_slice(&instance.mesh.indices),
                usage: wgpu::BufferUsages::INDEX,
            });

            self.model_mesh_cache.insert(
                key,
                CachedModelMesh {
                    vertex_buffer: vb,
                    index_buffer: ib,
                    index_count: instance.mesh.indices.len() as u32,
                },
            );
        }
    }

    /// Pre-build the model instance transform dynamic-UBO buffer and bind
    /// group.
    ///
    /// The buffer is only rebuilt when the transform fingerprint changes
    /// (instance positions, rotations, scales, or camera origin moved).
    /// On steady-state frames with a static model set and a stationary
    /// camera, this avoids the cost of `create_buffer_init` +
    /// `create_bind_group` entirely.
    fn cache_model_transforms(
        &mut self,
        device: &wgpu::Device,
        model_instances: &[ModelInstance],
        camera_origin: DVec3,
        state: &MapState,
    ) {
        let min_align = device.limits().min_uniform_buffer_offset_alignment as usize;
        let stride = 64_usize.div_ceil(min_align) * min_align;

        // Build a rolling fingerprint of the model transform inputs.
        let mut fp: u64 = model_instances.len() as u64;
        let origin_key = [
            (camera_origin.x * 100.0) as i64,
            (camera_origin.y * 100.0) as i64,
            (camera_origin.z * 100.0) as i64,
        ];
        fp = fp
            .wrapping_mul(31)
            .wrapping_add(origin_key[0] as u64)
            .wrapping_mul(31)
            .wrapping_add(origin_key[1] as u64)
            .wrapping_mul(31)
            .wrapping_add(origin_key[2] as u64);
        for instance in model_instances {
            fp = fp
                .wrapping_mul(31)
                .wrapping_add(instance.position.lat.to_bits())
                .wrapping_mul(31)
                .wrapping_add(instance.position.lon.to_bits())
                .wrapping_mul(31)
                .wrapping_add(instance.scale.to_bits())
                .wrapping_mul(31)
                .wrapping_add(instance.heading.to_bits());
        }

        if let Some(ref cached) = self.cached_model_transforms {
            if cached.fingerprint == fp && cached.instance_count == model_instances.len() {
                return;
            }
        }

        let mut transform_bytes = vec![0u8; stride * model_instances.len()];
        for (i, instance) in model_instances.iter().enumerate() {
            let terrain_elev = state.elevation_at(&instance.position);
            let altitude = instance.resolve_altitude(terrain_elev);

            let world_pos = state.camera().projection().project(&instance.position);
            let rel_x = (world_pos.position.x - camera_origin.x) as f32;
            let rel_y = (world_pos.position.y - camera_origin.y) as f32;
            let rel_z = (altitude - camera_origin.z) as f32;

            let scale = instance.scale as f32;
            let heading = instance.heading as f32;
            let pitch = instance.pitch as f32;
            let roll = instance.roll as f32;

            let rotation = glam::Quat::from_rotation_z(heading)
                * glam::Quat::from_rotation_x(pitch)
                * glam::Quat::from_rotation_y(roll);
            let transform = Mat4::from_translation(glam::Vec3::new(rel_x, rel_y, rel_z))
                * Mat4::from_quat(rotation)
                * Mat4::from_scale(glam::Vec3::splat(scale));

            let mat = transform.to_cols_array_2d();
            let mat_bytes = bytemuck::cast_slice(&mat);
            let offset = i * stride;
            transform_bytes[offset..offset + 64].copy_from_slice(mat_bytes);
        }

        let buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
            label: Some("model_transforms_cached_buf"),
            contents: &transform_bytes,
            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
        });

        let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("model_transforms_cached_bg"),
            layout: &self.model_pipeline.model_bind_group_layout,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: buffer.as_entire_binding(),
            }],
        });

        self.cached_model_transforms = Some(CachedModelTransforms {
            buffer,
            bind_group,
            stride,
            instance_count: model_instances.len(),
            fingerprint: fp,
        })
    }

    // -- Page bind group management ---------------------------------------

    /// Rebuild per-atlas-page bind groups when new atlas pages are created.
    fn rebuild_page_bind_groups(&mut self, device: &wgpu::Device) {
        let tile_pages = self.tile_atlas.page_count();
        while self.page_bind_groups.len() < tile_pages {
            let idx = self.page_bind_groups.len();
            let view = &self.tile_atlas.pages[idx].view;
            let bg = device.create_bind_group(&wgpu::BindGroupDescriptor {
                label: Some(&format!("rustial_tile_page_bg_{idx}")),
                layout: &self.tile_pipeline.texture_bind_group_layout,
                entries: &[
                    wgpu::BindGroupEntry {
                        binding: 0,
                        resource: wgpu::BindingResource::TextureView(view),
                    },
                    wgpu::BindGroupEntry {
                        binding: 1,
                        resource: wgpu::BindingResource::Sampler(&self.sampler),
                    },
                ],
            });
            self.page_bind_groups.push(bg);

            let terrain_bg = device.create_bind_group(&wgpu::BindGroupDescriptor {
                label: Some(&format!("rustial_terrain_page_bg_{idx}")),
                layout: &self.terrain_pipeline.texture_bind_group_layout,
                entries: &[
                    wgpu::BindGroupEntry {
                        binding: 0,
                        resource: wgpu::BindingResource::TextureView(view),
                    },
                    wgpu::BindGroupEntry {
                        binding: 1,
                        resource: wgpu::BindingResource::Sampler(&self.sampler),
                    },
                ],
            });
            self.page_terrain_bind_groups.push(terrain_bg);
        }

        let hillshade_pages = self.hillshade_atlas.page_count();
        while self.page_hillshade_bind_groups.len() < hillshade_pages {
            let idx = self.page_hillshade_bind_groups.len();
            let view = &self.hillshade_atlas.pages[idx].view;
            let bg = device.create_bind_group(&wgpu::BindGroupDescriptor {
                label: Some(&format!("rustial_hillshade_page_bg_{idx}")),
                layout: &self.hillshade_pipeline.texture_bind_group_layout,
                entries: &[
                    wgpu::BindGroupEntry {
                        binding: 0,
                        resource: wgpu::BindingResource::TextureView(view),
                    },
                    wgpu::BindGroupEntry {
                        binding: 1,
                        resource: wgpu::BindingResource::Sampler(&self.sampler),
                    },
                ],
            });
            self.page_hillshade_bind_groups.push(bg);
        }
    }

    // -- Shared-grid terrain rendering ------------------------------------

    fn get_or_create_shared_grid(
        &mut self,
        device: &wgpu::Device,
        resolution: u16,
    ) -> &SharedTerrainGridMesh {
        self.shared_terrain_grids
            .entry(resolution)
            .or_insert_with(|| {
                let (vertices, indices) = build_shared_terrain_grid(resolution as usize);
                let vb = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                    label: Some(&format!("terrain_grid_vb_{resolution}")),
                    contents: bytemuck::cast_slice(&vertices),
                    usage: wgpu::BufferUsages::VERTEX,
                });
                let ib = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
                    label: Some(&format!("terrain_grid_ib_{resolution}")),
                    contents: bytemuck::cast_slice(&indices),
                    usage: wgpu::BufferUsages::INDEX,
                });
                SharedTerrainGridMesh {
                    vertex_buffer: vb,
                    index_buffer: ib,
                    index_count: indices.len() as u32,
                }
            });
        #[allow(clippy::unwrap_used)] // Just inserted above.
        self.shared_terrain_grids.get(&resolution).unwrap()
    }

    fn get_or_create_terrain_tile_bind(
        &mut self,
        device: &wgpu::Device,
        queue: &wgpu::Queue,
        mesh: &TerrainMeshData,
        state: &MapState,
        scene_origin: DVec3,
        pipeline_kind: TerrainPipelineKind,
    ) -> Option<()> {
        let elevation = mesh.elevation_texture.as_ref()?;
        let key = TerrainTileBindKey {
            tile: mesh.tile,
            pipeline: pipeline_kind,
        };
        let origin_key = [
            (scene_origin.x * 100.0) as i64,
            (scene_origin.y * 100.0) as i64,
            (scene_origin.z * 100.0) as i64,
        ];

        // Check if already cached and valid.
        if let Some(cached) = self.terrain_tile_bind_cache.get(&key) {
            if cached.origin_key == origin_key && cached.generation == mesh.generation {
                return Some(());
            }
        }

        // Ensure height texture is cached.
        let tile = mesh.tile;
        let gen = mesh.generation;
        let needs_height = self
            .height_texture_cache
            .get(&tile)
            .is_none_or(|c| c.generation != gen);
        if needs_height {
            let size = wgpu::Extent3d {
                width: elevation.width.max(1),
                height: elevation.height.max(1),
                depth_or_array_layers: 1,
            };
            let texture = device.create_texture(&wgpu::TextureDescriptor {
                label: Some(&format!("height_tex_{:?}", tile)),
                size,
                mip_level_count: 1,
                sample_count: 1,
                dimension: wgpu::TextureDimension::D2,
                format: wgpu::TextureFormat::R32Float,
                usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::COPY_DST,
                view_formats: &[],
            });
            queue.write_texture(
                wgpu::TexelCopyTextureInfo {
                    texture: &texture,
                    mip_level: 0,
                    origin: wgpu::Origin3d::ZERO,
                    aspect: wgpu::TextureAspect::All,
                },
                bytemuck::cast_slice(&elevation.data),
                wgpu::TexelCopyBufferLayout {
                    offset: 0,
                    bytes_per_row: Some(elevation.width.max(1) * 4),
                    rows_per_image: None,
                },
                size,
            );
            let view = texture.create_view(&wgpu::TextureViewDescriptor::default());
            self.height_texture_cache.insert(
                tile,
                CachedHeightTexture {
                    generation: gen,
                    view,
                },
            );
        }

        // Build the tile uniform and bind group.
        let tile_uniform = build_terrain_tile_uniform(mesh, elevation, state, scene_origin);

        let uniform_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
            label: Some(&format!(
                "terrain_tile_uniform_{:?}_{:?}",
                mesh.tile, pipeline_kind
            )),
            contents: bytemuck::bytes_of(&tile_uniform),
            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
        });

        let layout = match pipeline_kind {
            TerrainPipelineKind::Terrain => &self.terrain_pipeline.tile_bind_group_layout,
            TerrainPipelineKind::TerrainData => &self.terrain_data_pipeline.tile_bind_group_layout,
            TerrainPipelineKind::Hillshade => &self.hillshade_pipeline.tile_bind_group_layout,
        };

        // Obtain the height texture view pointer safely.
        let height_view_ref = &self.height_texture_cache.get(&tile)?.view;
        let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some(&format!(
                "terrain_tile_bg_{:?}_{:?}",
                mesh.tile, pipeline_kind
            )),
            layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: uniform_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: wgpu::BindingResource::TextureView(height_view_ref),
                },
            ],
        });

        self.terrain_tile_bind_cache.insert(
            key,
            CachedTerrainTileBind {
                uniform_buffer,
                bind_group,
                origin_key,
                generation: mesh.generation,
            },
        );
        Some(())
    }

    fn render_shared_terrain_tiles<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        _state: &MapState,
        terrain_meshes: &[TerrainMeshData],
        visible_tiles: &[VisibleTile],
    ) {
        pass.set_pipeline(&self.terrain_pipeline.pipeline);
        pass.set_bind_group(0, &self.terrain_uniform_bind_group, &[]);

        for mesh in terrain_meshes {
            let grid = match self.shared_terrain_grids.get(&mesh.grid_resolution) {
                Some(g) => g,
                None => continue,
            };

            if let Some(actual) = find_terrain_texture_actual(mesh.tile, visible_tiles) {
                if let Some(region) = self.tile_atlas.get(&actual) {
                    if let Some(bg) = self.page_terrain_bind_groups.get(region.page) {
                        pass.set_bind_group(1, bg, &[]);
                    }
                }
            }

            let key = TerrainTileBindKey {
                tile: mesh.tile,
                pipeline: TerrainPipelineKind::Terrain,
            };
            if let Some(cached) = self.terrain_tile_bind_cache.get(&key) {
                pass.set_bind_group(2, &cached.bind_group, &[]);
                pass.set_vertex_buffer(0, grid.vertex_buffer.slice(..));
                pass.set_index_buffer(grid.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
                pass.draw_indexed(0..grid.index_count, 0, 0..1);
            }
        }
    }

    fn render_shared_terrain_data_tiles<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        _state: &MapState,
        terrain_meshes: &[TerrainMeshData],
    ) {
        pass.set_pipeline(&self.terrain_data_pipeline.pipeline);
        pass.set_bind_group(0, &self.terrain_data_uniform_bind_group, &[]);

        for mesh in terrain_meshes {
            let grid = match self.shared_terrain_grids.get(&mesh.grid_resolution) {
                Some(g) => g,
                None => continue,
            };

            let key = TerrainTileBindKey {
                tile: mesh.tile,
                pipeline: TerrainPipelineKind::TerrainData,
            };
            if let Some(cached) = self.terrain_tile_bind_cache.get(&key) {
                pass.set_bind_group(1, &cached.bind_group, &[]);
                pass.set_vertex_buffer(0, grid.vertex_buffer.slice(..));
                pass.set_index_buffer(grid.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
                pass.draw_indexed(0..grid.index_count, 0, 0..1);
            }
        }
    }

    fn render_shared_hillshade_tiles<'a>(
        &'a self,
        pass: &mut wgpu::RenderPass<'a>,
        _state: &MapState,
        terrain_meshes: &[TerrainMeshData],
    ) {
        pass.set_pipeline(&self.hillshade_pipeline.pipeline);
        pass.set_bind_group(0, &self.hillshade_uniform_bind_group, &[]);

        for mesh in terrain_meshes {
            let grid = match self.shared_terrain_grids.get(&mesh.grid_resolution) {
                Some(g) => g,
                None => continue,
            };

            if let Some(region) = self.hillshade_atlas.get(&mesh.tile) {
                if let Some(bg) = self.page_hillshade_bind_groups.get(region.page) {
                    pass.set_bind_group(1, bg, &[]);
                }
            }

            let key = TerrainTileBindKey {
                tile: mesh.tile,
                pipeline: TerrainPipelineKind::Hillshade,
            };
            if let Some(cached) = self.terrain_tile_bind_cache.get(&key) {
                pass.set_bind_group(2, &cached.bind_group, &[]);
                pass.set_vertex_buffer(0, grid.vertex_buffer.slice(..));
                pass.set_index_buffer(grid.index_buffer.slice(..), wgpu::IndexFormat::Uint32);
                pass.draw_indexed(0..grid.index_count, 0, 0..1);
            }
        }
    }

    // -- Cache pruning ----------------------------------------------------

    fn prune_height_texture_cache(&mut self, terrain_meshes: &[TerrainMeshData]) {
        let live: std::collections::HashSet<TileId> =
            terrain_meshes.iter().map(|m| m.tile).collect();
        self.height_texture_cache
            .retain(|tile, _| live.contains(tile));
    }

    fn prune_terrain_tile_bind_cache(&mut self, terrain_meshes: &[TerrainMeshData]) {
        let live: std::collections::HashSet<TileId> =
            terrain_meshes.iter().map(|m| m.tile).collect();
        self.terrain_tile_bind_cache
            .retain(|key, _| live.contains(&key.tile));
    }

    fn prune_grid_scalar_overlay_cache(&mut self, overlays: &[&VisualizationOverlay]) {
        let live: std::collections::HashSet<LayerId> = overlays
            .iter()
            .filter_map(|overlay| match overlay {
                VisualizationOverlay::GridScalar { layer_id, .. } => Some(*layer_id),
                _ => None,
            })
            .collect();
        self.grid_scalar_overlay_cache
            .retain(|layer_id, _| live.contains(layer_id));
    }

    fn prune_grid_extrusion_overlay_cache(&mut self, overlays: &[&VisualizationOverlay]) {
        let live: std::collections::HashSet<LayerId> = overlays
            .iter()
            .filter_map(|overlay| match overlay {
                VisualizationOverlay::GridExtrusion { layer_id, .. } => Some(*layer_id),
                _ => None,
            })
            .collect();
        self.grid_extrusion_overlay_cache
            .retain(|layer_id, _| live.contains(layer_id));
    }

    fn prune_column_overlay_cache(&mut self, overlays: &[&VisualizationOverlay]) {
        let live: std::collections::HashSet<LayerId> = overlays
            .iter()
            .filter_map(|overlay| match overlay {
                VisualizationOverlay::Columns { layer_id, .. } => Some(*layer_id),
                _ => None,
            })
            .collect();
        self.column_overlay_cache
            .retain(|layer_id, _| live.contains(layer_id));
    }

    fn prune_point_cloud_overlay_cache(&mut self, overlays: &[&VisualizationOverlay]) {
        let live: std::collections::HashSet<LayerId> = overlays
            .iter()
            .filter_map(|overlay| match overlay {
                VisualizationOverlay::Points { layer_id, .. } => Some(*layer_id),
                _ => None,
            })
            .collect();
        self.point_cloud_overlay_cache
            .retain(|layer_id, _| live.contains(layer_id));
    }

    // -- Headless capture -------------------------------------------------

    /// Render one full frame to an offscreen texture and return the pixel
    /// data as a `Vec<u8>` in RGBA8 format (4 bytes per pixel,
    /// `width * height * 4` total bytes).
    ///
    /// This is the primary entry-point for headless rendering and
    /// cross-renderer comparison tests.  It creates a transient
    /// colour texture, calls [`render_full`](Self::render_full), copies
    /// the result to a readback buffer, and returns the pixels.
    ///
    /// # Arguments
    ///
    /// * `state` - Engine map state (camera, layers, terrain).
    /// * `device` - WGPU device.
    /// * `queue` - WGPU queue.
    /// * `visible_tiles` - Tile set for this frame.
    /// * `vector_meshes` - Tessellated vector layers.
    /// * `model_instances` - 3D model instances.
    ///
    /// # Returns
    ///
    /// `Some(pixels)` on success, `None` if the GPU readback fails.
    pub fn render_to_buffer(
        &mut self,
        state: &MapState,
        device: &wgpu::Device,
        queue: &wgpu::Queue,
        visible_tiles: &[VisibleTile],
        vector_meshes: &[VectorMeshData],
        model_instances: &[ModelInstance],
    ) -> Option<Vec<u8>> {
        let format = wgpu::TextureFormat::Rgba8UnormSrgb;

        let color_tex = device.create_texture(&wgpu::TextureDescriptor {
            label: Some("rustial_render_to_buffer_color"),
            size: wgpu::Extent3d {
                width: self.width,
                height: self.height,
                depth_or_array_layers: 1,
            },
            mip_level_count: 1,
            sample_count: 1,
            dimension: wgpu::TextureDimension::D2,
            format,
            usage: wgpu::TextureUsages::RENDER_ATTACHMENT | wgpu::TextureUsages::COPY_SRC,
            view_formats: &[],
        });
        let color_view = color_tex.create_view(&wgpu::TextureViewDescriptor::default());

        let clear_color = state.computed_fog().clear_color;

        self.render_full(&RenderParams {
            state,
            device,
            queue,
            color_view: &color_view,
            visible_tiles,
            vector_meshes,
            model_instances,
            clear_color,
        });

        let bytes_per_row = self.width * 4;
        let buffer_size = (bytes_per_row * self.height) as wgpu::BufferAddress;
        let readback = device.create_buffer(&wgpu::BufferDescriptor {
            label: Some("rustial_render_to_buffer_readback"),
            size: buffer_size,
            usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::MAP_READ,
            mapped_at_creation: false,
        });

        let mut encoder = device.create_command_encoder(&wgpu::CommandEncoderDescriptor {
            label: Some("rustial_render_to_buffer_encoder"),
        });
        encoder.copy_texture_to_buffer(
            wgpu::TexelCopyTextureInfo {
                texture: &color_tex,
                mip_level: 0,
                origin: wgpu::Origin3d::ZERO,
                aspect: wgpu::TextureAspect::All,
            },
            wgpu::TexelCopyBufferInfo {
                buffer: &readback,
                layout: wgpu::TexelCopyBufferLayout {
                    offset: 0,
                    bytes_per_row: Some(bytes_per_row),
                    rows_per_image: Some(self.height),
                },
            },
            wgpu::Extent3d {
                width: self.width,
                height: self.height,
                depth_or_array_layers: 1,
            },
        );
        queue.submit(std::iter::once(encoder.finish()));

        let slice = readback.slice(..);
        let (tx, rx) = std::sync::mpsc::channel();
        slice.map_async(wgpu::MapMode::Read, move |res| {
            let _ = tx.send(res);
        });
        let _ = device.poll(wgpu::PollType::wait());
        rx.recv().ok()?.ok()?;

        let data = slice.get_mapped_range().to_vec();
        readback.unmap();
        Some(data)
    }

    /// Return the current renderer width in pixels.
    pub fn width(&self) -> u32 {
        self.width
    }

    /// Return the current renderer height in pixels.
    pub fn height(&self) -> u32 {
        self.height
    }

    /// Return per-frame visualization cache activity from the last render.
    pub fn visualization_perf_stats(&self) -> VisualizationPerfStats {
        self.visualization_perf_stats
    }

    /// Return atlas health diagnostics for the tile atlas.
    ///
    /// Useful for performance overlays and automated tests.  Call after
    /// [`render_full`](Self::render_full) for post-frame metrics, or at
    /// any time for the current snapshot.
    pub fn tile_atlas_diagnostics(&self) -> crate::gpu::tile_atlas::AtlasDiagnostics {
        self.tile_atlas.diagnostics()
    }

    /// Return atlas health diagnostics for the hillshade atlas.
    pub fn hillshade_atlas_diagnostics(&self) -> crate::gpu::tile_atlas::AtlasDiagnostics {
        self.hillshade_atlas.diagnostics()
    }
}

// ---------------------------------------------------------------------------
// Helpers (module-private)
// ---------------------------------------------------------------------------

fn build_grid_scalar_geometry(
    grid: &rustial_engine::GeoGrid,
    state: &MapState,
    scene_origin: DVec3,
) -> (Vec<GridScalarVertex>, Vec<u32>) {
    let rows = grid.rows.max(1);
    let cols = grid.cols.max(1);
    let mut vertices = Vec::with_capacity((rows + 1) * (cols + 1));
    let mut indices = Vec::with_capacity(rows * cols * 6);

    for row in 0..=rows {
        for col in 0..=cols {
            let u = col as f32 / cols as f32;
            let v = row as f32 / rows as f32;
            let coord = grid_corner_coord(grid, row, col, state);
            let projected = state.camera().projection().project(&coord);
            vertices.push(GridScalarVertex {
                position: [
                    (projected.position.x - scene_origin.x) as f32,
                    (projected.position.y - scene_origin.y) as f32,
                    (projected.position.z - scene_origin.z + 0.05) as f32,
                ],
                uv: [u, v],
            });
        }
    }

    for row in 0..rows {
        for col in 0..cols {
            let tl = (row * (cols + 1) + col) as u32;
            let tr = tl + 1;
            let bl = ((row + 1) * (cols + 1) + col) as u32;
            let br = bl + 1;
            indices.extend_from_slice(&[tl, bl, tr, tr, bl, br]);
        }
    }

    (vertices, indices)
}

fn grid_corner_coord(
    grid: &rustial_engine::GeoGrid,
    row: usize,
    col: usize,
    state: &MapState,
) -> rustial_math::GeoCoord {
    let dx = col as f64 * grid.cell_width;
    let dy = row as f64 * grid.cell_height;
    let (sin_r, cos_r) = grid.rotation.sin_cos();
    let rx = dx * cos_r - dy * sin_r;
    let ry = dx * sin_r + dy * cos_r;
    let coord = offset_geo_coord(&grid.origin, rx, ry);
    let altitude = resolve_grid_surface_altitude(grid, &coord, state);
    rustial_math::GeoCoord::new(coord.lat, coord.lon, altitude)
}

fn create_grid_scalar_texture(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
    field: &rustial_engine::ScalarField2D,
) -> wgpu::Texture {
    let size = wgpu::Extent3d {
        width: field.cols.max(1) as u32,
        height: field.rows.max(1) as u32,
        depth_or_array_layers: 1,
    };
    let texture = device.create_texture(&wgpu::TextureDescriptor {
        label: Some("grid_scalar_field_texture"),
        size,
        mip_level_count: 1,
        sample_count: 1,
        dimension: wgpu::TextureDimension::D2,
        format: wgpu::TextureFormat::R32Float,
        usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::COPY_DST,
        view_formats: &[],
    });
    write_grid_scalar_texture(queue, &texture, field);
    texture
}

fn write_grid_scalar_texture(
    queue: &wgpu::Queue,
    texture: &wgpu::Texture,
    field: &rustial_engine::ScalarField2D,
) {
    let size = wgpu::Extent3d {
        width: field.cols.max(1) as u32,
        height: field.rows.max(1) as u32,
        depth_or_array_layers: 1,
    };
    queue.write_texture(
        wgpu::TexelCopyTextureInfo {
            texture,
            mip_level: 0,
            origin: wgpu::Origin3d::ZERO,
            aspect: wgpu::TextureAspect::All,
        },
        bytemuck::cast_slice(&field.data),
        wgpu::TexelCopyBufferLayout {
            offset: 0,
            bytes_per_row: Some(field.cols.max(1) as u32 * 4),
            rows_per_image: Some(field.rows.max(1) as u32),
        },
        size,
    );
}

fn create_grid_scalar_ramp_texture(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
    ramp: &rustial_engine::ColorRamp,
) -> wgpu::Texture {
    let width = 256u32;
    let data = ramp.as_texture_data(width);
    let size = wgpu::Extent3d {
        width,
        height: 1,
        depth_or_array_layers: 1,
    };
    let texture = device.create_texture(&wgpu::TextureDescriptor {
        label: Some("grid_scalar_ramp_texture"),
        size,
        mip_level_count: 1,
        sample_count: 1,
        dimension: wgpu::TextureDimension::D2,
        format: wgpu::TextureFormat::Rgba8UnormSrgb,
        usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::COPY_DST,
        view_formats: &[],
    });
    queue.write_texture(
        wgpu::TexelCopyTextureInfo {
            texture: &texture,
            mip_level: 0,
            origin: wgpu::Origin3d::ZERO,
            aspect: wgpu::TextureAspect::All,
        },
        &data,
        wgpu::TexelCopyBufferLayout {
            offset: 0,
            bytes_per_row: Some(width * 4),
            rows_per_image: Some(1),
        },
        size,
    );
    texture
}

// ---------------------------------------------------------------------------
// Heatmap two-pass helpers
// ---------------------------------------------------------------------------

/// Create the off-screen R16Float accumulation texture for heatmap Pass 1.
fn create_heatmap_accum_texture(
    device: &wgpu::Device,
    width: u32,
    height: u32,
) -> (wgpu::Texture, wgpu::TextureView) {
    let texture = device.create_texture(&wgpu::TextureDescriptor {
        label: Some("heatmap_accum_texture"),
        size: wgpu::Extent3d {
            width,
            height,
            depth_or_array_layers: 1,
        },
        mip_level_count: 1,
        sample_count: 1,
        dimension: wgpu::TextureDimension::D2,
        format: wgpu::TextureFormat::R16Float,
        usage: wgpu::TextureUsages::RENDER_ATTACHMENT | wgpu::TextureUsages::TEXTURE_BINDING,
        view_formats: &[],
    });
    let view = texture.create_view(&wgpu::TextureViewDescriptor::default());
    (texture, view)
}

/// Create a default 256×1 Rgba8Unorm heatmap colour ramp texture.
///
/// The ramp interpolates through: transparent → royal blue → cyan → lime →
/// yellow → red, matching the MapLibre default heat stops.
#[allow(clippy::needless_range_loop)]
fn create_default_heatmap_ramp_texture(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
) -> wgpu::Texture {
    const WIDTH: u32 = 256;
    let stops: &[(f32, [u8; 4])] = &[
        (0.00, [0, 0, 0, 0]),
        (0.10, [65, 105, 225, 255]),
        (0.30, [0, 255, 255, 255]),
        (0.50, [0, 255, 0, 255]),
        (0.70, [255, 255, 0, 255]),
        (1.00, [255, 0, 0, 255]),
    ];

    let mut data = vec![0u8; WIDTH as usize * 4];
    for i in 0..WIDTH as usize {
        let t = i as f32 / (WIDTH - 1) as f32;
        // Find the two surrounding stops.
        let mut lo = 0;
        for s in 1..stops.len() {
            if stops[s].0 >= t {
                lo = s - 1;
                break;
            }
        }
        let hi = (lo + 1).min(stops.len() - 1);
        let range = stops[hi].0 - stops[lo].0;
        let frac = if range > 0.0 {
            (t - stops[lo].0) / range
        } else {
            0.0
        };
        for c in 0..4 {
            let a = stops[lo].1[c] as f32;
            let b = stops[hi].1[c] as f32;
            data[i * 4 + c] = (a + (b - a) * frac).round() as u8;
        }
    }

    let size = wgpu::Extent3d {
        width: WIDTH,
        height: 1,
        depth_or_array_layers: 1,
    };
    let texture = device.create_texture(&wgpu::TextureDescriptor {
        label: Some("heatmap_ramp_texture"),
        size,
        mip_level_count: 1,
        sample_count: 1,
        dimension: wgpu::TextureDimension::D2,
        format: wgpu::TextureFormat::Rgba8Unorm,
        usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::COPY_DST,
        view_formats: &[],
    });
    queue.write_texture(
        wgpu::TexelCopyTextureInfo {
            texture: &texture,
            mip_level: 0,
            origin: wgpu::Origin3d::ZERO,
            aspect: wgpu::TextureAspect::All,
        },
        &data,
        wgpu::TexelCopyBufferLayout {
            offset: 0,
            bytes_per_row: Some(WIDTH * 4),
            rows_per_image: Some(1),
        },
        size,
    );
    texture
}

/// Create the bind group for heatmap colour-mapping Pass 2 (group 1).
fn create_heatmap_colormap_bind_group(
    device: &wgpu::Device,
    layout: &wgpu::BindGroupLayout,
    accum_view: &wgpu::TextureView,
    ramp_view: &wgpu::TextureView,
    sampler: &wgpu::Sampler,
) -> wgpu::BindGroup {
    device.create_bind_group(&wgpu::BindGroupDescriptor {
        label: Some("heatmap_colormap_textures_bg"),
        layout,
        entries: &[
            wgpu::BindGroupEntry {
                binding: 0,
                resource: wgpu::BindingResource::TextureView(accum_view),
            },
            wgpu::BindGroupEntry {
                binding: 1,
                resource: wgpu::BindingResource::TextureView(ramp_view),
            },
            wgpu::BindGroupEntry {
                binding: 2,
                resource: wgpu::BindingResource::Sampler(sampler),
            },
        ],
    })
}

fn build_grid_scalar_uniform(
    grid: &rustial_engine::GeoGrid,
    field: &rustial_engine::ScalarField2D,
    state: &MapState,
    scene_origin: DVec3,
    opacity: f32,
) -> GridScalarUniform {
    let projection_kind = match state.camera().projection() {
        rustial_engine::CameraProjection::WebMercator => 0.0,
        rustial_engine::CameraProjection::Equirectangular => 1.0,
        _ => 0.0,
    };
    let base_altitude = match grid.altitude_mode {
        rustial_engine::AltitudeMode::ClampToGround => 0.0,
        rustial_engine::AltitudeMode::RelativeToGround => grid.origin.alt as f32,
        rustial_engine::AltitudeMode::Absolute => grid.origin.alt as f32,
    };
    GridScalarUniform {
        origin_counts: [
            grid.origin.lat as f32,
            grid.origin.lon as f32,
            grid.rows as f32,
            grid.cols as f32,
        ],
        grid_params: [
            grid.cell_width as f32,
            grid.cell_height as f32,
            grid.rotation as f32,
            opacity,
        ],
        scene_origin: [
            scene_origin.x as f32,
            scene_origin.y as f32,
            scene_origin.z as f32,
            projection_kind,
        ],
        value_params: [
            field.min,
            field.max,
            field.nan_value.unwrap_or(0.0),
            if field.nan_value.is_some() { 1.0 } else { 0.0 },
        ],
        base_altitude: [base_altitude, 0.0, 0.0, 0.0],
    }
}

fn grid_scalar_ramp_fingerprint(ramp: &rustial_engine::ColorRamp) -> u64 {
    let mut h = ramp.stops.len() as u64;
    for stop in &ramp.stops {
        h = h
            .wrapping_mul(31)
            .wrapping_add(stop.value.to_bits() as u64)
            .wrapping_mul(31)
            .wrapping_add(stop.color[0].to_bits() as u64)
            .wrapping_mul(31)
            .wrapping_add(stop.color[1].to_bits() as u64)
            .wrapping_mul(31)
            .wrapping_add(stop.color[2].to_bits() as u64)
            .wrapping_mul(31)
            .wrapping_add(stop.color[3].to_bits() as u64);
    }
    h
}

fn grid_extrusion_params_fingerprint(params: &rustial_engine::ExtrusionParams) -> u64 {
    (params.height_scale.to_bits())
        .wrapping_mul(31)
        .wrapping_add(params.base_meters.to_bits())
}

fn grid_extrusion_grid_fingerprint(grid: &rustial_engine::GeoGrid) -> u64 {
    let mut h = 17u64;
    h = h.wrapping_mul(31).wrapping_add(grid.origin.lat.to_bits());
    h = h.wrapping_mul(31).wrapping_add(grid.origin.lon.to_bits());
    h = h.wrapping_mul(31).wrapping_add(grid.origin.alt.to_bits());
    h = h.wrapping_mul(31).wrapping_add(grid.rows as u64);
    h = h.wrapping_mul(31).wrapping_add(grid.cols as u64);
    h = h.wrapping_mul(31).wrapping_add(grid.cell_width.to_bits());
    h = h.wrapping_mul(31).wrapping_add(grid.cell_height.to_bits());
    h = h.wrapping_mul(31).wrapping_add(grid.rotation.to_bits());
    h = h.wrapping_mul(31).wrapping_add(match grid.altitude_mode {
        rustial_engine::AltitudeMode::ClampToGround => 0,
        rustial_engine::AltitudeMode::RelativeToGround => 1,
        rustial_engine::AltitudeMode::Absolute => 2,
    });
    h
}

fn build_grid_extrusion_geometry(
    grid: &rustial_engine::GeoGrid,
    field: &rustial_engine::ScalarField2D,
    ramp: &rustial_engine::ColorRamp,
    params: &rustial_engine::ExtrusionParams,
    state: &MapState,
    scene_origin: DVec3,
) -> (Vec<GridExtrusionVertex>, Vec<u32>) {
    let mut vertices = Vec::new();
    let mut indices = Vec::new();

    for row in 0..grid.rows {
        for col in 0..grid.cols {
            let Some(value) = field.sample(row, col) else {
                continue;
            };

            let t = field.normalized(row, col).unwrap_or(0.5);
            let color = ramp.evaluate(t);
            let corners = grid_cell_corners_world(grid, row, col, state, scene_origin, params);
            append_extruded_cell_geometry(
                &mut vertices,
                &mut indices,
                corners,
                value * params.height_scale as f32,
                color,
            );
        }
    }

    (vertices, indices)
}

fn append_extruded_cell_geometry(
    vertices: &mut Vec<GridExtrusionVertex>,
    indices: &mut Vec<u32>,
    corners: [[f32; 3]; 4],
    extrusion_height: f32,
    color: [f32; 4],
) {
    let [nw, ne, sw, se] = corners;
    let top = [
        [nw[0], nw[1], nw[2] + extrusion_height],
        [ne[0], ne[1], ne[2] + extrusion_height],
        [sw[0], sw[1], sw[2] + extrusion_height],
        [se[0], se[1], se[2] + extrusion_height],
    ];
    let base = [nw, ne, sw, se];

    append_quad(
        vertices,
        indices,
        top[0],
        top[1],
        top[2],
        top[3],
        [0.0, 0.0, 1.0],
        color,
    );
    append_quad(
        vertices,
        indices,
        base[0],
        base[1],
        top[0],
        top[1],
        [0.0, -1.0, 0.0],
        color,
    );
    append_quad(
        vertices,
        indices,
        top[2],
        top[3],
        base[2],
        base[3],
        [0.0, 1.0, 0.0],
        color,
    );
    append_quad(
        vertices,
        indices,
        base[0],
        top[0],
        base[2],
        top[2],
        [-1.0, 0.0, 0.0],
        color,
    );
    append_quad(
        vertices,
        indices,
        top[1],
        base[1],
        top[3],
        base[3],
        [1.0, 0.0, 0.0],
        color,
    );
}

#[allow(clippy::too_many_arguments)]
fn append_quad(
    vertices: &mut Vec<GridExtrusionVertex>,
    indices: &mut Vec<u32>,
    a: [f32; 3],
    b: [f32; 3],
    c: [f32; 3],
    d: [f32; 3],
    normal: [f32; 3],
    color: [f32; 4],
) {
    let base_index = vertices.len() as u32;
    vertices.extend_from_slice(&[
        GridExtrusionVertex {
            position: a,
            normal,
            color,
        },
        GridExtrusionVertex {
            position: b,
            normal,
            color,
        },
        GridExtrusionVertex {
            position: c,
            normal,
            color,
        },
        GridExtrusionVertex {
            position: d,
            normal,
            color,
        },
    ]);
    indices.extend_from_slice(&[
        base_index,
        base_index + 2,
        base_index + 1,
        base_index + 1,
        base_index + 2,
        base_index + 3,
    ]);
}

fn grid_cell_corners_world(
    grid: &rustial_engine::GeoGrid,
    row: usize,
    col: usize,
    state: &MapState,
    scene_origin: DVec3,
    params: &rustial_engine::ExtrusionParams,
) -> [[f32; 3]; 4] {
    let nw = project_grid_offset(
        grid,
        col as f64 * grid.cell_width,
        row as f64 * grid.cell_height,
        state,
        scene_origin,
        params,
    );
    let ne = project_grid_offset(
        grid,
        (col + 1) as f64 * grid.cell_width,
        row as f64 * grid.cell_height,
        state,
        scene_origin,
        params,
    );
    let sw = project_grid_offset(
        grid,
        col as f64 * grid.cell_width,
        (row + 1) as f64 * grid.cell_height,
        state,
        scene_origin,
        params,
    );
    let se = project_grid_offset(
        grid,
        (col + 1) as f64 * grid.cell_width,
        (row + 1) as f64 * grid.cell_height,
        state,
        scene_origin,
        params,
    );
    [nw, ne, sw, se]
}

fn project_grid_offset(
    grid: &rustial_engine::GeoGrid,
    dx: f64,
    dy: f64,
    state: &MapState,
    scene_origin: DVec3,
    params: &rustial_engine::ExtrusionParams,
) -> [f32; 3] {
    let (sin_r, cos_r) = grid.rotation.sin_cos();
    let rx = dx * cos_r - dy * sin_r;
    let ry = dx * sin_r + dy * cos_r;
    let coord = offset_geo_coord(&grid.origin, rx, ry);
    let altitude = resolve_grid_base_altitude(grid, &coord, state, params) as f64;
    let elevated_coord = rustial_math::GeoCoord::new(coord.lat, coord.lon, altitude);
    let projected = state.camera().projection().project(&elevated_coord);
    [
        (projected.position.x - scene_origin.x) as f32,
        (projected.position.y - scene_origin.y) as f32,
        (projected.position.z - scene_origin.z) as f32,
    ]
}

fn offset_geo_coord(
    origin: &rustial_math::GeoCoord,
    dx_meters: f64,
    dy_meters: f64,
) -> rustial_math::GeoCoord {
    const METERS_PER_DEG_LAT: f64 = 111_320.0;
    let lat = origin.lat - dy_meters / METERS_PER_DEG_LAT;
    let cos_lat = origin.lat.to_radians().cos().max(1e-10);
    let lon = origin.lon + dx_meters / (METERS_PER_DEG_LAT * cos_lat);
    rustial_math::GeoCoord::new(lat, lon, origin.alt)
}

fn resolve_grid_base_altitude(
    grid: &rustial_engine::GeoGrid,
    coord: &rustial_math::GeoCoord,
    state: &MapState,
    params: &rustial_engine::ExtrusionParams,
) -> f32 {
    let terrain = state.elevation_at(coord).unwrap_or(0.0);
    match grid.altitude_mode {
        rustial_engine::AltitudeMode::ClampToGround => (terrain + params.base_meters) as f32,
        rustial_engine::AltitudeMode::RelativeToGround => {
            (terrain + grid.origin.alt + params.base_meters) as f32
        }
        rustial_engine::AltitudeMode::Absolute => (grid.origin.alt + params.base_meters) as f32,
    }
}

fn resolve_grid_surface_altitude(
    grid: &rustial_engine::GeoGrid,
    coord: &rustial_math::GeoCoord,
    state: &MapState,
) -> f64 {
    let terrain = state.elevation_at(coord).unwrap_or(0.0);
    match grid.altitude_mode {
        rustial_engine::AltitudeMode::ClampToGround => terrain,
        rustial_engine::AltitudeMode::RelativeToGround => terrain + grid.origin.alt,
        rustial_engine::AltitudeMode::Absolute => grid.origin.alt,
    }
}

fn build_unit_column_mesh() -> (Vec<ColumnVertex>, Vec<u32>) {
    let vertices = vec![
        // top
        ColumnVertex {
            position: [-0.5, -0.5, 1.0],
            normal: [0.0, 0.0, 1.0],
        },
        ColumnVertex {
            position: [0.5, -0.5, 1.0],
            normal: [0.0, 0.0, 1.0],
        },
        ColumnVertex {
            position: [-0.5, 0.5, 1.0],
            normal: [0.0, 0.0, 1.0],
        },
        ColumnVertex {
            position: [0.5, 0.5, 1.0],
            normal: [0.0, 0.0, 1.0],
        },
        // bottom
        ColumnVertex {
            position: [-0.5, -0.5, 0.0],
            normal: [0.0, 0.0, -1.0],
        },
        ColumnVertex {
            position: [0.5, -0.5, 0.0],
            normal: [0.0, 0.0, -1.0],
        },
        ColumnVertex {
            position: [-0.5, 0.5, 0.0],
            normal: [0.0, 0.0, -1.0],
        },
        ColumnVertex {
            position: [0.5, 0.5, 0.0],
            normal: [0.0, 0.0, -1.0],
        },
        // north
        ColumnVertex {
            position: [-0.5, -0.5, 0.0],
            normal: [0.0, -1.0, 0.0],
        },
        ColumnVertex {
            position: [0.5, -0.5, 0.0],
            normal: [0.0, -1.0, 0.0],
        },
        ColumnVertex {
            position: [-0.5, -0.5, 1.0],
            normal: [0.0, -1.0, 0.0],
        },
        ColumnVertex {
            position: [0.5, -0.5, 1.0],
            normal: [0.0, -1.0, 0.0],
        },
        // south
        ColumnVertex {
            position: [-0.5, 0.5, 1.0],
            normal: [0.0, 1.0, 0.0],
        },
        ColumnVertex {
            position: [0.5, 0.5, 1.0],
            normal: [0.0, 1.0, 0.0],
        },
        ColumnVertex {
            position: [-0.5, 0.5, 0.0],
            normal: [0.0, 1.0, 0.0],
        },
        ColumnVertex {
            position: [0.5, 0.5, 0.0],
            normal: [0.0, 1.0, 0.0],
        },
        // west
        ColumnVertex {
            position: [-0.5, -0.5, 0.0],
            normal: [-1.0, 0.0, 0.0],
        },
        ColumnVertex {
            position: [-0.5, -0.5, 1.0],
            normal: [-1.0, 0.0, 0.0],
        },
        ColumnVertex {
            position: [-0.5, 0.5, 0.0],
            normal: [-1.0, 0.0, 0.0],
        },
        ColumnVertex {
            position: [-0.5, 0.5, 1.0],
            normal: [-1.0, 0.0, 0.0],
        },
        // east
        ColumnVertex {
            position: [0.5, -0.5, 1.0],
            normal: [1.0, 0.0, 0.0],
        },
        ColumnVertex {
            position: [0.5, -0.5, 0.0],
            normal: [1.0, 0.0, 0.0],
        },
        ColumnVertex {
            position: [0.5, 0.5, 1.0],
            normal: [1.0, 0.0, 0.0],
        },
        ColumnVertex {
            position: [0.5, 0.5, 0.0],
            normal: [1.0, 0.0, 0.0],
        },
    ];
    let indices = vec![
        0, 2, 1, 1, 2, 3, 4, 5, 6, 5, 7, 6, 8, 10, 9, 9, 10, 11, 12, 14, 13, 13, 14, 15, 16, 18,
        17, 17, 18, 19, 20, 22, 21, 21, 22, 23,
    ];
    (vertices, indices)
}

fn build_column_instances(
    columns: &rustial_engine::ColumnInstanceSet,
    ramp: &rustial_engine::ColorRamp,
    state: &MapState,
    scene_origin: DVec3,
) -> Vec<ColumnInstanceData> {
    let (min_height, max_height) = column_height_range(columns);
    columns
        .columns
        .iter()
        .map(|column| {
            let projected = state.camera().projection().project(&column.position);
            let base_z = resolve_column_base_altitude(column, state);
            let normalized = if (max_height - min_height).abs() < f64::EPSILON {
                0.5
            } else {
                ((column.height - min_height) / (max_height - min_height)).clamp(0.0, 1.0)
            } as f32;
            let color = column.color.unwrap_or_else(|| ramp.evaluate(normalized));
            ColumnInstanceData {
                base_position: [
                    (projected.position.x - scene_origin.x) as f32,
                    (projected.position.y - scene_origin.y) as f32,
                    (base_z - scene_origin.z) as f32,
                ],
                dimensions: [column.width as f32, column.height as f32, 0.0, 0.0],
                color,
            }
        })
        .collect()
}

fn build_point_instances(
    points: &rustial_engine::PointInstanceSet,
    ramp: &rustial_engine::ColorRamp,
    state: &MapState,
    scene_origin: DVec3,
) -> Vec<ColumnInstanceData> {
    points
        .points
        .iter()
        .map(|point| {
            let projected = state.camera().projection().project(&point.position);
            let center_z = resolve_point_altitude(point, state);
            let diameter = (point.radius * 2.0) as f32;
            let color = point
                .color
                .unwrap_or_else(|| ramp.evaluate(point.intensity.clamp(0.0, 1.0)));
            ColumnInstanceData {
                base_position: [
                    (projected.position.x - scene_origin.x) as f32,
                    (projected.position.y - scene_origin.y) as f32,
                    (center_z - scene_origin.z - point.radius) as f32,
                ],
                dimensions: [diameter, diameter, 0.0, 0.0],
                color,
            }
        })
        .collect()
}

fn column_height_range(columns: &rustial_engine::ColumnInstanceSet) -> (f64, f64) {
    let mut min_height = f64::INFINITY;
    let mut max_height = f64::NEG_INFINITY;
    for column in &columns.columns {
        min_height = min_height.min(column.height);
        max_height = max_height.max(column.height);
    }
    if min_height.is_infinite() || max_height.is_infinite() {
        (0.0, 0.0)
    } else {
        (min_height, max_height)
    }
}

fn resolve_point_altitude(point: &rustial_engine::PointInstance, state: &MapState) -> f64 {
    let terrain = state.elevation_at(&point.position).unwrap_or(0.0);
    match point.altitude_mode {
        rustial_engine::AltitudeMode::ClampToGround => terrain,
        rustial_engine::AltitudeMode::RelativeToGround => terrain + point.position.alt,
        rustial_engine::AltitudeMode::Absolute => point.position.alt,
    }
}

fn resolve_column_base_altitude(column: &rustial_engine::ColumnInstance, state: &MapState) -> f64 {
    let terrain = state.elevation_at(&column.position).unwrap_or(0.0);
    match column.altitude_mode {
        rustial_engine::AltitudeMode::ClampToGround => terrain + column.base,
        rustial_engine::AltitudeMode::RelativeToGround => {
            terrain + column.position.alt + column.base
        }
        rustial_engine::AltitudeMode::Absolute => column.position.alt + column.base,
    }
}

fn column_set_fingerprint(columns: &rustial_engine::ColumnInstanceSet) -> u64 {
    let mut h = columns.columns.len() as u64;
    for column in &columns.columns {
        h = h
            .wrapping_mul(31)
            .wrapping_add(column.position.lat.to_bits());
        h = h
            .wrapping_mul(31)
            .wrapping_add(column.position.lon.to_bits());
        h = h
            .wrapping_mul(31)
            .wrapping_add(column.position.alt.to_bits());
        h = h.wrapping_mul(31).wrapping_add(column.height.to_bits());
        h = h.wrapping_mul(31).wrapping_add(column.base.to_bits());
        h = h.wrapping_mul(31).wrapping_add(column.width.to_bits());
        h = h.wrapping_mul(31).wrapping_add(column.pick_id);
        h = h.wrapping_mul(31).wrapping_add(match column.altitude_mode {
            rustial_engine::AltitudeMode::ClampToGround => 0,
            rustial_engine::AltitudeMode::RelativeToGround => 1,
            rustial_engine::AltitudeMode::Absolute => 2,
        });
        if let Some(color) = column.color {
            h = h.wrapping_mul(31).wrapping_add(color[0].to_bits() as u64);
            h = h.wrapping_mul(31).wrapping_add(color[1].to_bits() as u64);
            h = h.wrapping_mul(31).wrapping_add(color[2].to_bits() as u64);
            h = h.wrapping_mul(31).wrapping_add(color[3].to_bits() as u64);
        }
    }
    h
}

fn point_set_fingerprint(points: &rustial_engine::PointInstanceSet) -> u64 {
    let mut h = points.points.len() as u64;
    for point in &points.points {
        h = h
            .wrapping_mul(31)
            .wrapping_add(point.position.lat.to_bits());
        h = h
            .wrapping_mul(31)
            .wrapping_add(point.position.lon.to_bits());
        h = h
            .wrapping_mul(31)
            .wrapping_add(point.position.alt.to_bits());
        h = h.wrapping_mul(31).wrapping_add(point.radius.to_bits());
        h = h
            .wrapping_mul(31)
            .wrapping_add(point.intensity.to_bits() as u64);
        h = h.wrapping_mul(31).wrapping_add(point.pick_id);
        h = h.wrapping_mul(31).wrapping_add(match point.altitude_mode {
            rustial_engine::AltitudeMode::ClampToGround => 0,
            rustial_engine::AltitudeMode::RelativeToGround => 1,
            rustial_engine::AltitudeMode::Absolute => 2,
        });
        if let Some(color) = point.color {
            h = h.wrapping_mul(31).wrapping_add(color[0].to_bits() as u64);
            h = h.wrapping_mul(31).wrapping_add(color[1].to_bits() as u64);
            h = h.wrapping_mul(31).wrapping_add(color[2].to_bits() as u64);
            h = h.wrapping_mul(31).wrapping_add(color[3].to_bits() as u64);
        }
    }
    h
}

fn visualization_overlay_intersects_scene_viewport(
    overlay: &VisualizationOverlay,
    state: &MapState,
) -> bool {
    let scene_origin = state.scene_world_origin();
    let Some(bounds) = visualization_overlay_world_bounds(overlay, state, scene_origin) else {
        return false;
    };
    bounds.intersects(state.scene_viewport_bounds())
}

fn visualization_overlay_world_bounds(
    overlay: &VisualizationOverlay,
    state: &MapState,
    scene_origin: DVec3,
) -> Option<rustial_math::WorldBounds> {
    match overlay {
        VisualizationOverlay::GridScalar { grid, .. }
        | VisualizationOverlay::GridExtrusion { grid, .. } => {
            Some(grid_world_bounds(grid, state, scene_origin))
        }
        VisualizationOverlay::Columns { columns, .. } => {
            column_world_bounds(columns, state, scene_origin)
        }
        VisualizationOverlay::Points { points, .. } => {
            point_world_bounds(points, state, scene_origin)
        }
    }
}

fn grid_world_bounds(
    grid: &rustial_engine::GeoGrid,
    state: &MapState,
    scene_origin: DVec3,
) -> rustial_math::WorldBounds {
    let corners = [
        grid_corner_coord(grid, 0, 0, state),
        grid_corner_coord(grid, 0, grid.cols, state),
        grid_corner_coord(grid, grid.rows, 0, state),
        grid_corner_coord(grid, grid.rows, grid.cols, state),
    ];
    let projected: Vec<_> = corners
        .iter()
        .map(|coord| state.camera().projection().project(coord))
        .collect();
    let mut bounds = rustial_math::WorldBounds::new(
        rustial_math::WorldCoord::new(
            projected[0].position.x - scene_origin.x,
            projected[0].position.y - scene_origin.y,
            projected[0].position.z - scene_origin.z,
        ),
        rustial_math::WorldCoord::new(
            projected[0].position.x - scene_origin.x,
            projected[0].position.y - scene_origin.y,
            projected[0].position.z - scene_origin.z,
        ),
    );
    for projected in projected.into_iter().skip(1) {
        bounds.extend_point(&rustial_math::WorldCoord::new(
            projected.position.x - scene_origin.x,
            projected.position.y - scene_origin.y,
            projected.position.z - scene_origin.z,
        ));
    }
    bounds
}

fn point_world_bounds(
    points: &rustial_engine::PointInstanceSet,
    state: &MapState,
    scene_origin: DVec3,
) -> Option<rustial_math::WorldBounds> {
    let mut bounds: Option<rustial_math::WorldBounds> = None;
    for point in &points.points {
        let projected = state.camera().projection().project(&point.position);
        let radius = point.radius;
        let center_z = resolve_point_altitude(point, state) - scene_origin.z;
        let point_bounds = rustial_math::WorldBounds::new(
            rustial_math::WorldCoord::new(
                projected.position.x - scene_origin.x - radius,
                projected.position.y - scene_origin.y - radius,
                center_z - radius,
            ),
            rustial_math::WorldCoord::new(
                projected.position.x - scene_origin.x + radius,
                projected.position.y - scene_origin.y + radius,
                center_z + radius,
            ),
        );
        if let Some(existing) = bounds.as_mut() {
            existing.extend(&point_bounds);
        } else {
            bounds = Some(point_bounds);
        }
    }
    bounds
}

fn column_world_bounds(
    columns: &rustial_engine::ColumnInstanceSet,
    state: &MapState,
    scene_origin: DVec3,
) -> Option<rustial_math::WorldBounds> {
    let mut bounds: Option<rustial_math::WorldBounds> = None;
    for column in &columns.columns {
        let projected = state.camera().projection().project(&column.position);
        let base_z = resolve_column_base_altitude(column, state) - scene_origin.z;
        let half_width = column.width * 0.5;
        let column_bounds = rustial_math::WorldBounds::new(
            rustial_math::WorldCoord::new(
                projected.position.x - scene_origin.x - half_width,
                projected.position.y - scene_origin.y - half_width,
                base_z,
            ),
            rustial_math::WorldCoord::new(
                projected.position.x - scene_origin.x + half_width,
                projected.position.y - scene_origin.y + half_width,
                base_z + column.height,
            ),
        );
        if let Some(existing) = bounds.as_mut() {
            existing.extend(&column_bounds);
        } else {
            bounds = Some(column_bounds);
        }
    }
    bounds
}

fn build_shared_terrain_grid(resolution: usize) -> (Vec<TerrainGridVertex>, Vec<u32>) {
    let res = resolution.max(2);
    let mut vertices = Vec::with_capacity(res * res);
    let mut indices = Vec::with_capacity((res - 1) * (res - 1) * 6);

    for row in 0..res {
        for col in 0..res {
            let u = col as f32 / (res - 1) as f32;
            let v = row as f32 / (res - 1) as f32;
            vertices.push(TerrainGridVertex {
                uv: [u, v],
                skirt: 0.0,
            });
        }
    }

    for row in 0..(res - 1) {
        for col in 0..(res - 1) {
            let tl = (row * res + col) as u32;
            let tr = tl + 1;
            let bl = ((row + 1) * res + col) as u32;
            let br = bl + 1;
            indices.extend_from_slice(&[tl, bl, tr, tr, bl, br]);
        }
    }

    let edges: [Vec<usize>; 4] = [
        (0..res).collect(),
        ((res - 1) * res..res * res).collect(),
        (0..res).map(|r| r * res).collect(),
        (0..res).map(|r| r * res + res - 1).collect(),
    ];

    for edge in &edges {
        for i in 0..edge.len() - 1 {
            let a = edge[i] as u32;
            let b = edge[i + 1] as u32;
            let uv_a = vertices[edge[i]].uv;
            let uv_b = vertices[edge[i + 1]].uv;
            let base_a = vertices.len() as u32;
            let base_b = base_a + 1;
            vertices.push(TerrainGridVertex {
                uv: uv_a,
                skirt: 1.0,
            });
            vertices.push(TerrainGridVertex {
                uv: uv_b,
                skirt: 1.0,
            });
            indices.extend_from_slice(&[a, base_a, b, b, base_a, base_b]);
        }
    }

    (vertices, indices)
}

fn build_terrain_tile_uniform(
    mesh: &TerrainMeshData,
    elevation: &rustial_engine::TerrainElevationTexture,
    state: &MapState,
    scene_origin: DVec3,
) -> TerrainTileUniform {
    let nw = rustial_math::tile_to_geo(&mesh.tile);
    let se = rustial_math::tile_xy_to_geo(
        mesh.tile.zoom,
        mesh.tile.x as f64 + 1.0,
        mesh.tile.y as f64 + 1.0,
    );
    let projection_kind = match state.camera().projection() {
        rustial_engine::CameraProjection::WebMercator => 0.0,
        rustial_engine::CameraProjection::Equirectangular => 1.0,
        _ => 0.0,
    };
    let skirt =
        rustial_engine::skirt_height(mesh.tile.zoom, mesh.vertical_exaggeration as f64) as f32;
    let skirt_base = (elevation.min_elev * mesh.vertical_exaggeration - skirt).max(-skirt * 3.0);
    let elev_region = if mesh.tile != mesh.elevation_source_tile {
        rustial_engine::elevation_region_in_texture_space(
            mesh.elevation_region,
            elevation.width,
            elevation.height,
        )
    } else {
        mesh.elevation_region
    };
    TerrainTileUniform {
        geo_bounds: [nw.lat as f32, nw.lon as f32, se.lat as f32, se.lon as f32],
        scene_origin: [
            scene_origin.x as f32,
            scene_origin.y as f32,
            scene_origin.z as f32,
            projection_kind,
        ],
        elev_params: [
            mesh.vertical_exaggeration,
            skirt_base,
            elevation.min_elev,
            elevation.max_elev,
        ],
        elev_region: [
            elev_region.u_min,
            elev_region.v_min,
            elev_region.u_max,
            elev_region.v_max,
        ],
    }
}

fn build_model_vertices(mesh: &rustial_engine::ModelMesh) -> Vec<ModelVertex> {
    debug_assert_eq!(mesh.positions.len(), mesh.normals.len());
    debug_assert_eq!(mesh.positions.len(), mesh.uvs.len());

    mesh.positions
        .iter()
        .zip(mesh.normals.iter())
        .zip(mesh.uvs.iter())
        .map(|((pos, normal), uv)| ModelVertex {
            position: *pos,
            normal: *normal,
            uv: *uv,
        })
        .collect()
}

#[cfg(test)]
mod tests {
    use super::*;
    use rustial_engine::{
        ColorRamp, ColorStop, ColumnInstance, ColumnInstanceSet, GeoCoord, GeoGrid,
        VisualizationOverlay,
    };

    fn visible_tile_with_fade(fade_opacity: f32) -> VisibleTile {
        let id = TileId::new(3, 4, 2);
        VisibleTile {
            target: id,
            actual: id,
            data: None,
            fade_opacity,
        }
    }

    fn test_ramp() -> ColorRamp {
        ColorRamp::new(vec![
            ColorStop {
                value: 0.0,
                color: [0.0, 0.0, 1.0, 0.5],
            },
            ColorStop {
                value: 1.0,
                color: [1.0, 0.0, 0.0, 0.8],
            },
        ])
    }

    #[test]
    fn tile_batch_cache_key_changes_when_fade_opacity_changes() {
        let a = [visible_tile_with_fade(0.25)];
        let b = [visible_tile_with_fade(0.75)];

        let key_a = TileBatchCacheKey::new(
            &a,
            DVec3::ZERO,
            rustial_engine::CameraProjection::WebMercator,
        );
        let key_b = TileBatchCacheKey::new(
            &b,
            DVec3::ZERO,
            rustial_engine::CameraProjection::WebMercator,
        );

        assert_ne!(
            key_a, key_b,
            "tile batch cache key must include fade-sensitive inputs"
        );
    }

    #[test]
    fn tile_batch_cache_key_stays_equal_when_fade_opacity_matches() {
        let a = [visible_tile_with_fade(1.0)];
        let b = [visible_tile_with_fade(1.0)];

        let key_a = TileBatchCacheKey::new(
            &a,
            DVec3::ZERO,
            rustial_engine::CameraProjection::WebMercator,
        );
        let key_b = TileBatchCacheKey::new(
            &b,
            DVec3::ZERO,
            rustial_engine::CameraProjection::WebMercator,
        );

        assert_eq!(key_a, key_b);
    }

    #[test]
    fn diff_column_instance_ranges_tracks_contiguous_changes() {
        let old = vec![
            ColumnInstanceData {
                base_position: [0.0, 0.0, 0.0],
                dimensions: [1.0, 2.0, 0.0, 0.0],
                color: [1.0, 0.0, 0.0, 1.0],
            },
            ColumnInstanceData {
                base_position: [1.0, 0.0, 0.0],
                dimensions: [1.0, 2.0, 0.0, 0.0],
                color: [0.0, 1.0, 0.0, 1.0],
            },
            ColumnInstanceData {
                base_position: [2.0, 0.0, 0.0],
                dimensions: [1.0, 2.0, 0.0, 0.0],
                color: [0.0, 0.0, 1.0, 1.0],
            },
            ColumnInstanceData {
                base_position: [3.0, 0.0, 0.0],
                dimensions: [1.0, 2.0, 0.0, 0.0],
                color: [1.0, 1.0, 0.0, 1.0],
            },
        ];
        let mut new = old.clone();
        new[1].dimensions[1] = 4.0;
        new[2].color = [1.0, 0.0, 1.0, 1.0];

        assert_eq!(diff_column_instance_ranges(&old, &new), vec![1..3]);
    }

    #[test]
    fn diff_column_instance_ranges_splits_disjoint_changes() {
        let old = vec![
            ColumnInstanceData {
                base_position: [0.0, 0.0, 0.0],
                dimensions: [1.0, 2.0, 0.0, 0.0],
                color: [1.0, 0.0, 0.0, 1.0],
            },
            ColumnInstanceData {
                base_position: [1.0, 0.0, 0.0],
                dimensions: [1.0, 2.0, 0.0, 0.0],
                color: [0.0, 1.0, 0.0, 1.0],
            },
            ColumnInstanceData {
                base_position: [2.0, 0.0, 0.0],
                dimensions: [1.0, 2.0, 0.0, 0.0],
                color: [0.0, 0.0, 1.0, 1.0],
            },
        ];
        let mut new = old.clone();
        new[0].dimensions[0] = 3.0;
        new[2].base_position[2] = 5.0;

        assert_eq!(diff_column_instance_ranges(&old, &new), vec![0..1, 2..3]);
    }

    #[test]
    fn visualization_overlay_visibility_rejects_far_grid() {
        let mut state = MapState::new();
        state.set_viewport(1280, 720);
        state.set_camera_target(GeoCoord::from_lat_lon(0.0, 0.0));
        state.set_camera_distance(1_000.0);
        state.update_camera(1.0 / 60.0);

        let overlay = VisualizationOverlay::GridScalar {
            layer_id: LayerId::next(),
            grid: GeoGrid::new(GeoCoord::from_lat_lon(70.0, 120.0), 2, 2, 50.0, 50.0),
            field: rustial_engine::ScalarField2D::from_data(2, 2, vec![1.0; 4]),
            ramp: test_ramp(),
        };

        assert!(!visualization_overlay_intersects_scene_viewport(
            &overlay, &state
        ));
    }

    #[test]
    fn visualization_overlay_visibility_accepts_near_columns() {
        let mut state = MapState::new();
        state.set_viewport(1280, 720);
        state.set_camera_target(GeoCoord::from_lat_lon(0.0, 0.0));
        state.set_camera_distance(1_000.0);
        state.update_camera(1.0 / 60.0);

        let overlay = VisualizationOverlay::Columns {
            layer_id: LayerId::next(),
            columns: ColumnInstanceSet::new(vec![ColumnInstance::new(
                GeoCoord::from_lat_lon(0.0, 0.0),
                10.0,
                5.0,
            )]),
            ramp: test_ramp(),
        };

        assert!(visualization_overlay_intersects_scene_viewport(
            &overlay, &state
        ));
    }
}

// ---------------------------------------------------------------------------
// Shadow map texture helpers
// ---------------------------------------------------------------------------

/// Create `cascade_count` Depth32Float textures (one per shadow cascade) and
/// their views.
fn create_shadow_map_textures(
    device: &wgpu::Device,
    resolution: u32,
    cascade_count: u32,
) -> (Vec<wgpu::Texture>, Vec<wgpu::TextureView>) {
    let mut textures = Vec::with_capacity(cascade_count as usize);
    let mut views = Vec::with_capacity(cascade_count as usize);
    for i in 0..cascade_count {
        let tex = device.create_texture(&wgpu::TextureDescriptor {
            label: Some(&format!("shadow_cascade_{i}")),
            size: wgpu::Extent3d {
                width: resolution,
                height: resolution,
                depth_or_array_layers: 1,
            },
            mip_level_count: 1,
            sample_count: 1,
            dimension: wgpu::TextureDimension::D2,
            format: wgpu::TextureFormat::Depth32Float,
            usage: wgpu::TextureUsages::RENDER_ATTACHMENT | wgpu::TextureUsages::TEXTURE_BINDING,
            view_formats: &[],
        });
        let view = tex.create_view(&wgpu::TextureViewDescriptor::default());
        textures.push(tex);
        views.push(view);
    }
    (textures, views)
}

/// GPU-side shadow parameters uniform (matches WGSL `ShadowParams`).
///
/// Layout: 2 × mat4x4 (128 B) + 2 × vec4 (32 B) = 160 B total.
#[repr(C)]
#[derive(Debug, Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
struct ShadowParamsUniform {
    /// Light-space VP matrix for cascade 0.
    light_matrix_0: [[f32; 4]; 4],
    /// Light-space VP matrix for cascade 1.
    light_matrix_1: [[f32; 4]; 4],
    /// `[intensity, texel_size, normal_offset, cascade_split]`
    shadow_config: [f32; 4],
    /// `[dir.x, dir.y, dir.z, enabled (0 or 1)]`
    shadow_dir: [f32; 4],
}