roxlap_render/

lib.rs

//! roxlap-render — unified CPU/GPU renderer facade.
//!
//! One [`SceneRenderer`] hides the choice between the CPU opticast
//! path (`roxlap-core` / `roxlap-scene`, presented via `softbuffer`)
//! and the GPU compute-shader path (`roxlap-gpu`, presented via its
//! own wgpu surface). Construction picks the GPU backend when asked
//! and able, and **falls back to CPU automatically** when WGPU init
//! fails — so a host never has to branch on GPU availability or carry
//! the `Scene`→GPU upload/refresh/transform glue itself.
//!
//! Hosts stay thin: build a `Scene`, advance it from input, then call
//! [`SceneRenderer::render`] each frame. The facade owns the window
//! surface, the framebuffer/z-buffer (CPU) or the resident scene +
//! dirty-chunk tracking (GPU), and presentation.
//!
//! The per-frame flow is `render` → *(optional overlays)* → finish.
//! Between [`SceneRenderer::render`] and the finishing
//! [`SceneRenderer::present`] / [`SceneRenderer::paint_egui`] call, a
//! host may overlay depth-tested world-space lines with
//! [`SceneRenderer::draw_lines`] (editor gizmos, debug geometry — see
//! [`Line3`]); they land in the framebuffer, occluded by the rendered
//! scene, with egui still painting panels on top.
//!
//! This is the RF.0 skeleton: backend selection + fallback + a
//! clear-to-sky frame. RF.1/RF.2 fill in the real CPU/GPU scene
//! render; RF.3 adds sprites; RF.4 adds framebuffer capture.

#![forbid(unsafe_code)]

mod cpu;
/// WebGL2 framebuffer presenter for the CPU backend on wasm (the
/// browser has no `softbuffer`).
#[cfg(target_arch = "wasm32")]
mod cpu_blit;
#[cfg(feature = "hud")]
mod cpu_egui;
mod gpu;

#[cfg(not(target_arch = "wasm32"))]
use std::sync::Arc;

use roxlap_core::opticast::OpticastSettings;
use roxlap_core::sky::Sky;
use roxlap_core::sprite::SpriteLighting;
use roxlap_core::Camera;
use roxlap_scene::Scene;

pub use roxlap_formats::kfa::KfaSprite;
pub use roxlap_formats::sprite::Sprite;
pub use roxlap_gpu::{GpuInitError, GpuRendererSettings, PowerPreference};
// Re-exported so hosts can name the [`SceneRenderer::new`] bounds
// without adding a direct `raw-window-handle` dependency of their own.
pub use raw_window_handle::{HasDisplayHandle, HasWindowHandle};
// Re-exported so hosts feed [`SceneRenderer::paint_egui`] from the exact
// egui version the renderer was built against (`hud` feature).
#[cfg(feature = "hud")]
pub use egui;

use crate::cpu::CpuBackend;
use crate::gpu::GpuBackend;

/// Type-erased display handle stored by the CPU backend's softbuffer
/// surface. `raw-window-handle` implements `HasDisplayHandle` for
/// `Arc<H>` (`H: ?Sized`), and the bare trait object implements its
/// own object-safe trait — so `Arc<W>` coerces to `Arc<DynDisplay>`
/// for any provider `W`.
#[cfg(not(target_arch = "wasm32"))]
pub(crate) type DynDisplay = dyn HasDisplayHandle + Send + Sync + 'static;
/// Type-erased window handle counterpart to [`DynDisplay`].
#[cfg(not(target_arch = "wasm32"))]
pub(crate) type DynWindow = dyn HasWindowHandle + Send + Sync + 'static;

/// One placed sprite instance: which [`SpriteSet::models`] entry and
/// where in the world.
pub struct SpriteInstanceDesc {
    pub model: usize,
    pub pos: [f32; 3],
}

/// Backend-agnostic sprite description. The facade builds the CPU
/// per-instance draw list and the GPU instanced registry from the
/// same data, so both backends show identical sprites. The host owns
/// content (which models, where, recolouring) — building a recoloured
/// variant is just a second [`Sprite`] model with edited `kv6.voxels`.
pub struct SpriteSet {
    /// Distinct voxel models (KV6 + base orientation). Instances index
    /// into this; their position overrides the model's.
    pub models: Vec<Sprite>,
    pub instances: Vec<SpriteInstanceDesc>,
    /// Model the [`SceneRenderer::carve_active_sprite`] hotkey edits
    /// (GPU only, mirroring the demo's `G`-carve). `None` disables it.
    pub carve_model: Option<usize>,
}

/// Per-frame inputs both backends consume. The host builds the
/// [`OpticastSettings`] (it owns scan distance etc.); the facade does
/// everything else (pool config, sky fill, render, present).
pub struct FrameParams<'a> {
    /// CPU opticast settings (scan distance, mip ladder, framebuffer
    /// geometry). Ignored by the GPU backend.
    pub settings: &'a OpticastSettings,
    /// Packed engine sky colour: the CPU sky-miss fill + skycast, and
    /// the clear colour if no scene renders.
    pub sky_color: u32,
    /// Optional sky panorama for the CPU rasterizer's sky sampling.
    pub sky: Option<&'a Sky>,
    /// CPU fog: packed colour + max scan distance (voxels). `0` scan
    /// distance disables CPU fog.
    pub fog_color: u32,
    pub fog_max_scan_dist: i32,
    /// CPU: treat z=255 as air (avoids the S1.X bedrock path for
    /// out-of-bounds cameras).
    pub treat_z_max_as_air: bool,
    /// GPU scene-grid LOD scan distance (world units); see GPU.11.1.
    /// Ignored by the CPU backend.
    pub gpu_mip_scan_dist: f32,
    /// GPU outer-DDA step budget (chunks). Ignored by the CPU backend.
    pub gpu_max_outer_steps: u32,
    /// GPU vertical field of view (radians). Ignored by the CPU
    /// backend (it derives projection from [`OpticastSettings`]).
    pub gpu_fov_y_rad: f32,
    /// CPU sprite shading (built by the host from its engine). Required
    /// for the CPU backend to draw sprites; ignored by the GPU backend
    /// (its sprite pass shades from the uploaded model colours). `None`
    /// skips CPU sprite drawing.
    pub sprite_lighting: Option<&'a SpriteLighting<'a>>,
    /// Per-face directional shading for the voxel grids — voxlap's
    /// `setsideshades(top, bot, left, right, up, down)`, the grid-scan
    /// analogue of [`sprite_lighting`](Self::sprite_lighting). Each
    /// entry darkens the faces pointing that way; the host typically
    /// passes its engine's `side_shades()`. The default `[0; 6]` keeps
    /// `sideshademode` off (no per-side shading), so existing hosts and
    /// the oracle goldens are unaffected. Applied each frame by **both**
    /// backends: the CPU rasteriser via `gcsub`, and the GPU scene-DDA
    /// pass by darkening a hit voxel's brightness by the hit face's
    /// shade (the face taken from the DDA's last-stepped axis).
    pub side_shades: [i8; 6],
}

/// Result of [`SceneRenderer::pick`] — a resolved screen→world voxel
/// hit. `world` is the surface point (`cam.pos + t · normalize(ray)`);
/// `grid` + `voxel` are the owning grid and its **grid-local** voxel
/// (transform-correct for rotated / translated grids).
#[derive(Clone, Copy, PartialEq, Debug)]
pub struct PickHit {
    pub world: [f32; 3],
    pub grid: roxlap_scene::GridId,
    pub voxel: glam::IVec3,
}

/// A world-space view ray: the canonical unproject output of
/// [`SceneRenderer::view_ray`]. `dir` is unit-length. Feed it straight
/// to [`roxlap_scene::Scene::raycast`] for depth-free, backend-agnostic
/// voxel picking (`scene.raycast(ray.origin, ray.dir, max_dist)`), or
/// intersect it with a plane for tile selection.
#[derive(Clone, Copy, PartialEq, Debug)]
pub struct Ray {
    pub origin: glam::DVec3,
    pub dir: glam::DVec3,
}

/// A world-space line segment to draw over a rendered frame via
/// [`SceneRenderer::draw_lines`] — editor gizmos (bounding boxes, floor
/// grids, axes, hover wireframes), debug paths, etc.
#[derive(Clone, Copy, PartialEq, Debug)]
pub struct Line3 {
    /// World-space endpoints (voxel units), in the same frame the
    /// rendered scene + `camera` use.
    pub a: [f64; 3],
    pub b: [f64; 3],
    /// `0xAARRGGBB` — the high byte is an alpha blend factor (`0xFF`
    /// opaque, `0x00` invisible), the low 24 bits the RGB colour.
    pub color: u32,
    /// Screen-space thickness in pixels (`<= 1.0` draws a 1px line).
    pub width_px: f32,
    /// `true`: the segment is occluded by nearer rendered geometry
    /// (depth-tested against the frame's z-buffer). `false`: always on
    /// top (e.g. a hover highlight that should show through the model).
    pub depth_test: bool,
}

/// Which renderer a [`SceneRenderer`] resolved to at construction.
#[derive(Clone, Copy, PartialEq, Eq, Debug)]
pub enum Backend {
    /// `roxlap-core` opticast, presented via `softbuffer`.
    Cpu,
    /// `roxlap-gpu` compute marcher, presented via wgpu.
    Gpu,
}

/// Construction-time options for [`SceneRenderer::new`].
pub struct RenderOptions {
    /// Try the GPU backend first. When `false`, or when GPU init
    /// fails, the renderer uses the CPU backend.
    pub want_gpu: bool,
    /// Settings forwarded to [`roxlap_gpu::GpuRenderer`] when the GPU
    /// backend is selected.
    pub gpu: GpuRendererSettings,
    /// Packed `0x00RRGGBB` (alpha ignored) the empty/clear frame fills
    /// with until a scene render lands. Also the CPU sky-miss colour
    /// default if a frame supplies none.
    pub clear_sky: u32,
    /// CPU [`ScratchPool`](roxlap_core::rasterizer::ScratchPool) `lastx`
    /// sizing — the largest combined grid `vsid` the CPU rasterizer
    /// will see. Pre-sizing keeps later frames allocation-free.
    pub cpu_max_grid_vsid: u32,
    /// CPU strip-parallel render thread count (capped to the rayon
    /// pool). One [`ScratchPool`](roxlap_core::rasterizer::ScratchPool)
    /// slot per thread.
    pub cpu_render_threads: usize,
}

impl Default for RenderOptions {
    fn default() -> Self {
        Self {
            want_gpu: false,
            gpu: GpuRendererSettings::default(),
            clear_sky: 0x0099_b3d9,
            // 32 chunks × CHUNK_SIZE_XY — the scene-demo's widest
            // combined ground grid.
            cpu_max_grid_vsid: 32 * roxlap_scene::CHUNK_SIZE_XY,
            cpu_render_threads: 4,
        }
    }
}

/// Renderer-internal backend; never exposes wgpu or softbuffer types.
/// The GPU variant owns the whole wgpu device/queue/pipelines, so
/// it's boxed to keep the enum small.
enum BackendImpl {
    // Both variants boxed so the enum stays small regardless of which
    // backend's state is larger (clippy::large_enum_variant).
    Cpu(Box<CpuBackend>),
    Gpu(Box<GpuBackend>),
}

/// Unified renderer over the CPU and GPU paths. See the crate docs.
pub struct SceneRenderer {
    inner: BackendImpl,
}

impl SceneRenderer {
    /// Build a renderer for `window` — any [`raw-window-handle`]
    /// provider (winit, SDL, GLFW, …) in an `Arc`. `size` is the
    /// window's initial physical framebuffer size in pixels; thereafter
    /// the host reports changes via [`Self::resize`]. Passing the size
    /// explicitly keeps the facade decoupled from any one windowing
    /// library's size API.
    ///
    /// Selects the GPU backend when `opts.want_gpu` and WGPU
    /// initialises; otherwise the CPU backend. **Never fails** — a
    /// missing/incompatible GPU silently yields the CPU path (the
    /// message is logged to stderr).
    ///
    /// [`raw-window-handle`]: raw_window_handle
    #[cfg(not(target_arch = "wasm32"))]
    #[must_use]
    pub fn new<W>(window: Arc<W>, size: (u32, u32), opts: &RenderOptions) -> Self
    where
        W: HasWindowHandle + HasDisplayHandle + Send + Sync + 'static,
    {
        if opts.want_gpu {
            match GpuBackend::new(window.clone(), size, opts) {
                Ok(g) => {
                    return Self {
                        inner: BackendImpl::Gpu(Box::new(g)),
                    };
                }
                Err(e) => {
                    eprintln!(
                        "roxlap-render: GPU init failed ({e}); falling back to the CPU renderer",
                    );
                }
            }
        }
        Self {
            inner: BackendImpl::Cpu(Box::new(CpuBackend::new(window, size, opts))),
        }
    }

    /// wasm/WebGPU build-time entry: build a renderer over an HTML
    /// `canvas`. `size` is the canvas's initial framebuffer size in
    /// pixels; the host reports later changes via [`Self::resize`].
    ///
    /// Async because the browser drives wgpu's adapter/device requests
    /// through its event loop — `await` it inside a
    /// `wasm_bindgen_futures::spawn_local` task. Selects the GPU
    /// (WebGPU) backend when `opts.want_gpu` and WebGPU is available;
    /// otherwise (no WebGPU, or init failed) it falls back to the CPU
    /// opticast path presented through a WebGL2 blit on the same canvas.
    /// **Never fails** — the message is logged to the browser console.
    #[cfg(target_arch = "wasm32")]
    pub async fn new_from_canvas_async(
        canvas: web_sys::HtmlCanvasElement,
        size: (u32, u32),
        opts: &RenderOptions,
    ) -> Self {
        if opts.want_gpu {
            // `SurfaceTarget::Canvas` moves the canvas into wgpu, so the
            // GPU attempt gets a clone — the CPU fallback keeps the
            // original if WebGPU init fails.
            match GpuBackend::new_async(canvas.clone(), size, opts).await {
                Ok(g) => {
                    return Self {
                        inner: BackendImpl::Gpu(Box::new(g)),
                    };
                }
                Err(e) => {
                    web_sys::console::warn_1(
                        &format!("roxlap-render: WebGPU init failed ({e}); using the CPU renderer")
                            .into(),
                    );
                }
            }
        }
        Self {
            inner: BackendImpl::Cpu(Box::new(CpuBackend::new_from_canvas(canvas, size, opts))),
        }
    }

    /// Which backend was selected.
    #[must_use]
    pub fn backend(&self) -> Backend {
        match self.inner {
            BackendImpl::Cpu(_) => Backend::Cpu,
            BackendImpl::Gpu(_) => Backend::Gpu,
        }
    }

    /// The GPU adapter description when on the GPU backend, else
    /// `None`.
    #[must_use]
    pub fn adapter_info(&self) -> Option<&str> {
        match &self.inner {
            BackendImpl::Gpu(g) => Some(g.adapter_info()),
            BackendImpl::Cpu(_) => None,
        }
    }

    /// Upload an equirectangular sky panorama (RGBA8, `w×h`) for the
    /// GPU marcher's sky sampling. No-op on the CPU backend, which
    /// samples the [`Sky`] passed in each [`FrameParams`] instead.
    pub fn set_sky_panorama(&mut self, rgba: &[u8], w: u32, h: u32) {
        if let BackendImpl::Gpu(g) = &mut self.inner {
            g.set_sky_panorama(rgba, w, h);
        }
    }

    /// Follow a window resize. CPU resizes its framebuffer lazily, so
    /// this only matters to the GPU swapchain — but it's safe to call
    /// for both.
    pub fn resize(&mut self, width: u32, height: u32) {
        match &mut self.inner {
            BackendImpl::Cpu(c) => c.resize(width, height),
            BackendImpl::Gpu(g) => g.resize(width, height),
        }
    }

    /// Composite `scene` from `camera` with `frame` params into the
    /// backend's frame buffer — **without presenting**. The CPU backend
    /// fills sky + runs the opticast compositor into an owned buffer;
    /// the GPU backend uploads/refreshes the scene, runs the compute
    /// marcher + sprite pass, and acquires (but does not present) the
    /// swapchain frame.
    ///
    /// Finish the frame with exactly one of [`present`](Self::present)
    /// (no overlay) or [`paint_egui`](Self::paint_egui) (UI overlay).
    /// Calling `render` again without finishing drops the pending frame.
    pub fn render(&mut self, scene: &mut Scene, camera: &Camera, frame: &FrameParams) {
        match &mut self.inner {
            BackendImpl::Cpu(c) => c.render(scene, camera, frame),
            BackendImpl::Gpu(g) => g.render(scene, camera, frame),
        }
    }

    /// Draw world-space [`Line3`] segments over the frame
    /// [`render`](Self::render) composited, using that frame's camera +
    /// projection + depth buffer. Call **after** [`render`](Self::render)
    /// and **before** [`present`](Self::present) /
    /// [`paint_egui`](Self::paint_egui) — the lines land in the
    /// framebuffer, so a subsequent `paint_egui` still draws its panels
    /// on top.
    ///
    /// `camera` must be the one the last frame rendered with (the
    /// projection is taken from that frame). Depth-tested segments
    /// (`Line3::depth_test`) are occluded by nearer rendered geometry;
    /// always-on-top segments ignore depth. See [`Line3`] for colour /
    /// width / blend semantics.
    pub fn draw_lines(&mut self, camera: &Camera, lines: &[Line3]) {
        match &mut self.inner {
            BackendImpl::Cpu(c) => c.draw_lines(camera, lines),
            BackendImpl::Gpu(g) => g.draw_lines(camera, lines),
        }
    }

    /// Present the frame [`render`](Self::render) composited, with no UI
    /// overlay. Pairs with `render`; use [`paint_egui`](Self::paint_egui)
    /// instead to overlay an egui UI before presenting.
    pub fn present(&mut self) {
        match &mut self.inner {
            BackendImpl::Cpu(c) => c.present(),
            BackendImpl::Gpu(g) => g.present(),
        }
    }

    /// Overlay an egui UI on the frame [`render`](Self::render)
    /// composited, then present it (`hud` feature). The host runs egui
    /// itself (e.g. `egui` + `egui-winit`) and passes the tessellated
    /// `jobs` ([`egui::Context::tessellate`]) and the per-frame
    /// `textures` delta from [`egui::FullOutput`]; `pixels_per_point` is
    /// the UI scale (`ctx.pixels_per_point()`).
    ///
    /// The GPU backend paints via `egui-wgpu`; the CPU backend
    /// software-rasterises the tessellation into its framebuffer. Use
    /// this **instead of** [`present`](Self::present) — both finish the
    /// frame.
    #[cfg(feature = "hud")]
    pub fn paint_egui(
        &mut self,
        jobs: &[egui::ClippedPrimitive],
        textures: &egui::TexturesDelta,
        pixels_per_point: f32,
    ) {
        match &mut self.inner {
            BackendImpl::Cpu(c) => c.paint_egui(jobs, textures, pixels_per_point),
            BackendImpl::Gpu(g) => g.paint_egui(jobs, textures, pixels_per_point),
        }
    }

    /// Register sprite models + instances. The CPU backend builds a
    /// per-instance draw list; the GPU backend builds an instanced
    /// model registry. Call once at setup (or again to replace).
    pub fn set_sprites(&mut self, set: &SpriteSet) {
        match &mut self.inner {
            BackendImpl::Cpu(c) => c.set_sprites(set),
            BackendImpl::Gpu(g) => g.set_sprites(set),
        }
    }

    /// Register animated KFA sprites (one or more bone hierarchies).
    /// The GPU backend uploads each limb's kv6 as an instanced model
    /// **once** (appended to the sprite registry) and seeds the limb
    /// instances at their current pose; the CPU backend caches the
    /// posed limbs for drawing. Call once at setup, after
    /// [`set_sprites`](Self::set_sprites), then drive motion per frame
    /// with [`update_kfa_poses`](Self::update_kfa_poses).
    ///
    /// Limbs are posed from the sprites' current
    /// [`kfaval`](roxlap_formats::kfa::KfaSprite::kfaval) (advance
    /// [`animsprite`](roxlap_formats::kfa::KfaSprite::animsprite) first
    /// if using a baked curve), so `kfas` is taken `&mut`.
    pub fn set_kfa_sprites(&mut self, kfas: &mut [KfaSprite]) {
        match &mut self.inner {
            BackendImpl::Cpu(c) => c.set_kfa_sprites(kfas),
            BackendImpl::Gpu(g) => g.set_kfa_sprites(kfas),
        }
    }

    /// Re-pose the registered KFA sprites from their current
    /// `kfaval[]`. Call each frame after advancing the animation
    /// (`kfa.animsprite(dt_ms)` or poking `kfaval[]`). The GPU backend
    /// takes the cheap transform-only update (no model-volume
    /// re-upload); the CPU backend re-solves limb transforms for the
    /// next [`render`](Self::render). Must follow a
    /// [`set_kfa_sprites`](Self::set_kfa_sprites) with the same sprites.
    pub fn update_kfa_poses(&mut self, kfas: &mut [KfaSprite]) {
        match &mut self.inner {
            BackendImpl::Cpu(c) => c.update_kfa_poses(kfas),
            BackendImpl::Gpu(g) => g.update_kfa_poses(kfas),
        }
    }

    /// Carve the next z-layer off the [`SpriteSet::carve_model`] and
    /// re-upload (the demo's `G` hotkey + GPU.12 copy-on-modify). GPU
    /// only; a no-op on the CPU backend. Returns the voxels removed.
    pub fn carve_active_sprite(&mut self) -> u32 {
        match &mut self.inner {
            BackendImpl::Cpu(_) => 0,
            BackendImpl::Gpu(g) => g.carve_active_sprite(),
        }
    }

    /// Request that the next [`render`](Self::render) capture its
    /// framebuffer for [`take_capture`](Self::take_capture). CPU only
    /// (the GPU swapchain isn't read back) — a no-op on GPU.
    pub fn request_capture(&mut self) {
        if let BackendImpl::Cpu(c) = &mut self.inner {
            c.request_capture();
        }
    }

    /// Take the most recently captured frame as packed `0x00RRGGBB`
    /// pixels + dimensions, or `None` if no capture is ready / GPU.
    pub fn take_capture(&mut self) -> Option<(Vec<u32>, u32, u32)> {
        match &mut self.inner {
            BackendImpl::Cpu(c) => c.take_capture(),
            BackendImpl::Gpu(_) => None,
        }
    }

    /// Screen→world picking input: the world-space hit distance `t` at
    /// window pixel `(x, y)` from the **last rendered frame**, or `None`
    /// for out-of-bounds pixels and sky / no-hit. The host reconstructs
    /// the world hit point as `cam.pos + t * normalize(ray_dir)`, where
    /// `ray_dir` is the same per-pixel ray the frame was rendered with
    /// (see the backend's projection).
    ///
    /// `t` is the distance to the nearest **scene-grid** surface
    /// (terrain + grids); sprites do not occlude it (the sprite pass
    /// reads depth read-only), so a cursor sprite under the pointer is
    /// transparent to the pick.
    ///
    /// Cost: the CPU backend reads its in-memory z-buffer (free); the
    /// GPU backend stages the depth buffer and blocks on a device poll
    /// (cheap at click time — do not call every frame). The GPU path
    /// only has depth when the last frame drew sprites (`write_depth`).
    #[must_use]
    pub fn pick_depth(&self, x: u32, y: u32) -> Option<f32> {
        match &self.inner {
            BackendImpl::Cpu(c) => c.pick_depth(x, y),
            BackendImpl::Gpu(g) => g.pick_depth(x, y),
        }
    }

    /// World-space view-ray direction (un-normalised) for window pixel
    /// `(x, y)`, under the projection the **last frame** rendered with.
    /// The backends differ (CPU `setcamera` vs GPU vertical-FOV
    /// pinhole), so this hides which one is active. `None` before the
    /// first frame. Intersect it with a plane for tile picking, or feed
    /// it to [`Self::pick`] for a voxel.
    #[must_use]
    pub fn pixel_ray(&self, camera: &Camera, x: f64, y: f64) -> Option<[f64; 3]> {
        match &self.inner {
            BackendImpl::Cpu(c) => c.pixel_ray(camera, x, y),
            BackendImpl::Gpu(g) => g.pixel_ray(camera, x, y),
        }
    }

    /// Canonical screen→world unproject: the full view [`Ray`]
    /// (`camera.pos` origin + unit direction) for window pixel
    /// `(x, y)`, under whichever projection the last frame used. The
    /// one entry point both backends honour — hosts never reconstruct
    /// the projection. `None` before the first frame or for a
    /// degenerate ray.
    ///
    /// Compose with [`roxlap_scene::Scene::raycast`] for depth-free
    /// picking that's identical on CPU and GPU:
    /// `renderer.view_ray(cam, x, y).and_then(|r| scene.raycast(r.origin, r.dir, max))`.
    #[must_use]
    pub fn view_ray(&self, camera: &Camera, x: f64, y: f64) -> Option<Ray> {
        let d = self.pixel_ray(camera, x, y)?;
        let len = (d[0] * d[0] + d[1] * d[1] + d[2] * d[2]).sqrt();
        if len < 1e-12 {
            return None;
        }
        Some(Ray {
            origin: glam::DVec3::from_array([camera.pos[0], camera.pos[1], camera.pos[2]]),
            dir: glam::DVec3::new(d[0] / len, d[1] / len, d[2] / len),
        })
    }

    /// One-call screen→world voxel pick: unproject pixel `(x, y)` with
    /// the active backend's projection, read the last frame's depth
    /// there, reconstruct the world hit, and resolve it to the owning
    /// grid + grid-local voxel via [`Scene::resolve_voxel`]. `None` on
    /// sky / no-hit, or when no grid claims the surface.
    ///
    /// `scene` and `camera` must be the ones the last frame rendered;
    /// the projection (size + FOV / `hx,hy,hz`) is taken from that
    /// frame. Cheap on CPU (in-memory z-buffer); on GPU it stages the
    /// depth buffer (a click-time device poll — not per frame).
    #[must_use]
    pub fn pick(&self, scene: &Scene, camera: &Camera, x: u32, y: u32) -> Option<PickHit> {
        let dir = self.pixel_ray(camera, f64::from(x), f64::from(y))?;
        let t = f64::from(self.pick_depth(x, y)?);
        let len = (dir[0] * dir[0] + dir[1] * dir[1] + dir[2] * dir[2]).sqrt();
        if len < 1e-9 {
            return None;
        }
        let s = t / len; // world = cam.pos + t · (dir / |dir|)
        let world = glam::DVec3::new(
            camera.pos[0] + dir[0] * s,
            camera.pos[1] + dir[1] * s,
            camera.pos[2] + dir[2] * s,
        );
        let (grid, voxel) = scene.resolve_voxel(world, glam::DVec3::from_array(dir))?;
        #[allow(clippy::cast_possible_truncation)]
        let world_f32 = [world.x as f32, world.y as f32, world.z as f32];
        Some(PickHit {
            world: world_f32,
            grid,
            voxel,
        })
    }
}

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

    #[test]
    fn options_default_is_cpu_intent() {
        let o = RenderOptions::default();
        assert!(!o.want_gpu);
        assert_eq!(o.clear_sky & 0xFF00_0000, 0, "clear_sky is 0x00RRGGBB");
    }
}