concinnity_core/render/

backend.rs

//! RenderBackend trait: the union of methods every graphics backend
//! implements, dispatched dynamically by GraphicsSystem so the per-frame
//! step + setup logic lives in one cfg-free copy instead of three.
//!
//! Each concrete backend (MtlContext / DxContext / VkContext) supplies a
//! thin forwarder impl that delegates to the existing inherent methods
//! (see metal/backend.rs, directx/backend.rs, vulkan/backend.rs).
//!
//! Two cross-backend signature variances are handled here:
//!   - `upload_skinned`: Metal uses three shader payloads (vert + frag +
//!     shadow); DX/VK use one (frag). The trait method takes all three;
//!     DX/VK ignore the unused bytes.
//!   - `setup_chunk_streaming`: Metal binds chunk textures per draw and
//!     ignores the (texture_slot, normal_map_slot) args; DX/VK bake them
//!     into a shared descriptor at setup time.
//!
//! `render_stats` has a default no-op impl so a backend with no draw-call /
//! object counters need not override it; all three shipping backends do.

use crate::gfx::auto_exposure::AutoExposureSettings;
use crate::gfx::mesh_payload::{SkinnedVertex, Vertex};
use crate::gfx::profile::RenderStats;
use crate::gfx::render_types::{
    LineVertex, MaterialUniforms, PostProcessTunables, SkinnedDrawObject, TextDrawCall,
};
use crate::gfx::rt_reflections::RtReflectionSettings;
use crate::gfx::ssao::SsaoSettings;
use crate::gfx::ssgi::SsgiSettings;
use crate::gfx::ssr::SsrSettings;
use crate::render::backend_init::{BackendInit, ShaderBytes, SwapchainConfig};
use crate::render::error::{RenderError, RenderResult};
use crate::render::input::RenderInput;
use crate::render::keymap::KeyMap;
use crate::render::scene_flow::SceneControl;
use crate::render::volumetric_fog::FogSettings;
use alloc::string::String;
use alloc::string::ToString;
use alloc::vec::Vec;

/// Per-frame inputs for [`RenderBackend::draw_frame`]. `world_hidden` is set when
/// an opaque menu backdrop covers the scene: the backend skips every world pass
/// and presents only the overlay (`text_calls`) over a cleared target.
#[derive(Clone, Copy)]
pub struct FrameParams<'a> {
    /// Seconds since the world started, for time-driven effects.
    pub elapsed: f32,
    /// Vertical field of view in radians.
    pub fov_y_radians: f32,
    /// Near clip distance in world units.
    pub near: f32,
    /// Far clip distance in world units.
    pub far: f32,
    /// World-space camera position.
    pub cam_pos: [f32; 3],
    /// Overlay draw calls for this frame.
    pub text_calls: &'a [TextDrawCall],
    /// Expanded line ribbons (`lines::build_vertices`) for this frame's camera,
    /// drawn depth-tested into the scene after the world passes. Empty on any
    /// frame that submits no lines, which also drops the pass from the graph.
    pub lines: &'a [LineVertex],
    /// `true` when an opaque menu backdrop covers the scene.
    pub world_hidden: bool,
    /// Viewport view mode + show flags for the frame (`ViewOverrides` when the
    /// editor publishes one, defaults otherwise). Backends run their seeded
    /// graph inputs through `render_graph::apply_view` and steer the composite
    /// by the mode.
    pub view_mode: crate::gfx::view_modes::ViewMode,
    /// Feature passes to run this frame.
    pub show: crate::gfx::view_modes::ShowFlags,
}

/// One streamed chunk's geometry plus placement, supplied to
/// [`RenderBackend::add_chunk_mesh`]. `frame` reclaims retired deferred frees
/// before the chunk is placed in the streaming headroom.
#[derive(Clone, Copy)]
pub struct ChunkMesh<'a> {
    /// Chunk vertices.
    pub verts: &'a [Vertex],
    /// Chunk indices, mesh-relative.
    pub idxs: &'a [u16],
    /// Column-major placement matrix.
    pub model: [[f32; 4]; 4],
    /// Index into the shared texture pool for the albedo map.
    pub texture_slot: usize,
    /// Index into the shared texture pool for the normal map.
    pub normal_map_slot: usize,
    /// Per-chunk material scalars.
    pub material: MaterialUniforms,
    /// Current frame number, used to reclaim retired deferred frees.
    pub frame: u64,
}

/// One draw slot's fresh geometry, supplied to
/// [`RenderBackend::rebuild_static_geometry`] when an asset hot-reload
/// changed its vertex / index count and the slot can no longer hold the new
/// data in place. The backend rebuilds the entire shared vertex / index
/// buffer; draws not named here keep their current geometry, copied byte-for-
/// byte from the live buffers. `indices` are mesh-relative (0-based); the
/// backend rebases them onto whatever new vertex region the draw lands in.
pub struct DrawGeometryUpdate {
    /// The draw slot whose geometry is replaced.
    pub draw_idx: usize,
    /// Replacement vertices.
    pub vertices: Vec<Vertex>,
    /// Replacement indices, mesh-relative.
    pub indices: Vec<u16>,
    /// One slice per additional LOD, ordered mip 0 → mip N-1. Each is
    /// `(switch_distance, mesh-relative indices)`. Empty for meshes
    /// declared `lod_levels <= 1`.
    pub lod_alternates: Vec<(f32, Vec<u16>)>,
}

/// One skinned draw slot's fresh geometry, supplied to
/// [`RenderBackend::rebuild_skinned_geometry`] when an asset hot-reload
/// changed its vertex / index count and the slot can no longer hold the new
/// data in its existing region of the shared skinned vertex / index buffers.
/// The backend rebuilds both shared buffers; slots not named here keep their
/// current geometry, copied byte-for-byte from the live buffers and re-based
/// onto whatever new vertex region they land in. `indices` are mesh-relative
/// (0-based); the backend rebases them onto the new vertex region.
pub struct SkinnedDrawGeometryUpdate {
    /// The skinned slot whose geometry is replaced.
    pub skinned_index: usize,
    /// Replacement vertices.
    pub vertices: Vec<SkinnedVertex>,
    /// Replacement indices, mesh-relative.
    pub indices: Vec<u16>,
}

/// The post-rebuild layout for one skinned slot, returned by
/// [`RenderBackend::rebuild_skinned_geometry`] so the asset hot-reload
/// helper can refresh its `SkinnedMeshSourceEntry`s'
/// `vertex_base` / `vertex_count` / `index_count` to point at the new
/// regions. Returned for every slot (both the ones whose geometry was
/// replaced and the ones whose geometry was carried over) because the
/// rebuild may have shifted every slot's `vertex_base`.
/// Constructed only by the `cn debug` binary's skinned-rebuild reload pass;
/// reads as dead under `cargo check --lib`.
pub struct SkinnedSlotLayout {
    /// The skinned slot this layout describes.
    pub skinned_index: usize,
    /// First vertex of the slot's region in the shared skinned buffer.
    pub vertex_base: u32,
    /// Vertices in the slot's region.
    pub vertex_count: usize,
    /// Indices in the slot's region.
    pub index_count: usize,
}

/// The resolved per-feature quality settings for [`RenderBackend::apply_quality_settings`].
/// `GraphicsSystem` derives these from its stored `PostProcessConfig` (with the
/// user's persisted toggle overrides applied) whenever a Quality-group toggle
/// changes, so the backend receives ready-to-use settings rather than re-deriving
/// from the asset. Each `Option` mirrors the init-time gate: `None` means the
/// feature is off and its passes / resources should be torn down; `Some` means it
/// is on and its resources should exist. A backend without a live-rebuild path
/// ignores this (the choice still persists and applies at the next launch).
pub struct QualitySettings {
    /// Temporal anti-aliasing on/off (the `Taa` anti-aliasing mode). The backend
    /// additionally suppresses TAA while temporal upscaling is active (the scaler
    /// does its own accumulation). The other anti-aliasing modes are the composite
    /// FXAA edge filter, which rides `PostProcessTunables.fxaa` (pushed via
    /// `update_post_process`), not this pass-rebuild payload.
    pub taa: bool,
    /// Screen-space ambient occlusion, or `None` when off.
    pub ssao: Option<SsaoSettings>,
    /// Screen-space reflections, or `None` when off.
    pub ssr: Option<SsrSettings>,
    /// Hardware ray-traced reflections. The backend further gates this on GPU
    /// ray-tracing support, falling back to leaving it off when unsupported.
    pub rt_reflections: Option<RtReflectionSettings>,
    /// Screen-space global illumination, or `None` when off.
    pub ssgi: Option<SsgiSettings>,
    /// Per-axis divisor for the roughness-aware reflection blur target (the
    /// reduced-resolution first pass of the SSR / RT reflection composite),
    /// resolved from `PostProcessConfig.reflection_blur_resolution`. Every backend
    /// sizes its blur target at render / this on a live reflection rebuild.
    pub reflection_blur_scale: u32,
    /// Auto-exposure, or `None` when off.
    pub auto_exposure: Option<AutoExposureSettings>,
    /// The authored exposure bias (stops) auto-exposure applies on top of its
    /// adapted value; carried so a live auto-exposure enable matches init.
    pub auto_exposure_bias_ev: f32,
}

/// GPU/device capability flags, queried from the backend once it is built.
/// Surfaced so the settings menu can gray out (and make inert) toggles the
/// device cannot honor -- e.g. ray-traced reflections on a GPU without hardware
/// ray tracing. Mirrors an RHI-style capability set: a handful of bools held in
/// memory and re-queried each launch, never persisted, so it is always correct
/// for the current device + driver.
#[derive(Clone, Copy, Debug)]
pub struct DeviceCapabilities {
    /// Hardware ray tracing for the RT-reflections pass: DXR 1.1 on DirectX, the
    /// ray-query device extensions on Vulkan (and not under XeSS), and
    /// `MTLDevice::supportsRaytracing` on Metal.
    pub ray_tracing: bool,
    /// Whether the upscaler implementation is a choice (FSR3 / DLSS / XeSS)
    /// rather than fixed. DirectX and Vulkan offer the selection; Metal always
    /// upscales through MetalFX, so the row has nothing to pick.
    pub selectable_upscaler: bool,
    /// Whether a retired build-time draw slot may be recycled by a runtime
    /// clone. Metal's per-frame RT topology refresh re-admits recycled
    /// build-time slots; DirectX / Vulkan key their cull BVH + RT tables to
    /// fixed build-time indices and cannot refit, so only the runtime-append
    /// region recycles there. Read by the engine's draw-slot allocator.
    pub reuses_build_slots: bool,
    /// Whether a built draw slot's material and cull distance may be rewritten
    /// in place ([`RenderBackend::set_draw_material`] /
    /// [`RenderBackend::set_draw_cull_distance`]). Metal rebuilds its per-object
    /// buffer from the draw list every frame, so a rewritten slot draws with the
    /// new material next frame; DirectX / Vulkan bake per-object material state
    /// at build time and would keep drawing the old one. Read by the editor's
    /// live draw seam, which sends the edit to a world rebuild instead.
    pub rewrites_draws: bool,
}

impl DeviceCapabilities {
    /// Every capability present. The trait default, so a backend that does not
    /// report capabilities never wrongly disables a toggle (it keeps the prior
    /// behavior: the feature no-ops with a warning on an incapable device).
    pub const ALL: Self = Self {
        ray_tracing: true,
        selectable_upscaler: true,
        reuses_build_slots: true,
        rewrites_draws: true,
    };
}

impl Default for DeviceCapabilities {
    fn default() -> Self {
        Self::ALL
    }
}

/// Coarse GPU vendor class, derived per backend from the adapter's reported
/// vendor id (DirectX / Vulkan) or unified-memory / Apple-family signals (Metal).
/// Used only to pick default quality and to gate vendor-specific options (e.g.
/// which upscalers to offer); never persisted.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum GpuVendor {
    /// Apple silicon.
    Apple,
    /// NVIDIA.
    Nvidia,
    /// AMD.
    Amd,
    /// Intel.
    Intel,
    /// A vendor the probe does not recognise.
    Other,
}

/// Coarse performance class for default-quality selection, ordered low -> high so
/// callers can compare with `>=`. Each backend maps its native signals (memory
/// budget, discrete / integrated, Apple GPU family) onto this via `classify_tier`.
#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord)]
pub enum GpuTier {
    /// Unknown hardware: the conservative default, never the top preset. Sorts
    /// lowest so a comparison-based resolver treats it as the floor.
    Unknown,
    /// Integrated / low-power GPU: the lowest quality tier.
    Integrated,
    /// Older or small discrete GPU, or an Apple base M-series: entry quality.
    EntryDiscrete,
    /// Mainstream discrete GPU, or an Apple Pro: mid quality.
    MidDiscrete,
    /// Enthusiast discrete GPU, or an Apple Max / Ultra: high quality.
    HighDiscrete,
}

/// A coarse, Copy snapshot of the active GPU's class, queried from the backend
/// once it is built (mirrors `DeviceCapabilities`). Read at init to choose
/// sensible default graphics quality; never persisted, re-queried each launch so
/// it is always correct for the current device + driver. The GPU *name* is
/// deliberately omitted (it is not `Copy`); a backend exposes the name separately
/// when a UI needs it.
#[derive(Clone, Copy, Debug)]
pub struct GpuProfile {
    /// The GPU's vendor.
    pub vendor: GpuVendor,
    /// The performance tier the probe placed the GPU in.
    pub tier: GpuTier,
    /// Dedicated VRAM on a discrete GPU, or the recommended working-set on a
    /// unified-memory GPU. 0 when the backend / driver cannot report it.
    pub memory_budget_bytes: u64,
    /// Whether the GPU shares memory with the host.
    pub unified_memory: bool,
    /// Whether the GPU is a discrete card.
    pub discrete: bool,
}

impl GpuProfile {
    /// Conservative fallback for a backend that does not report a profile:
    /// unknown hardware picks the cautious baseline, never a high preset. The
    /// opposite default from `DeviceCapabilities::ALL` -- a feature gate fails
    /// open (assume capable, no-op with a warning if not), but quality
    /// auto-config fails safe (assume modest, never overdrive a weak GPU).
    pub const UNKNOWN: Self = Self {
        vendor: GpuVendor::Other,
        tier: GpuTier::Unknown,
        memory_budget_bytes: 0,
        unified_memory: false,
        discrete: false,
    };
}

impl Default for GpuProfile {
    fn default() -> Self {
        Self::UNKNOWN
    }
}

/// The cheap signals every backend can gather about its GPU, mapped to a coarse
/// `GpuTier` by one shared rule so the three backends classify consistently and
/// the mapping is unit-testable without a GPU. The backends differ in what they
/// can report (Apple exposes a GPU family; DirectX / Vulkan expose a VRAM figure
/// and a discrete / integrated flag), so this carries the union and the rule
/// uses whichever signals are present.
pub struct GpuClassInput {
    /// The GPU's vendor.
    pub vendor: GpuVendor,
    /// Device memory the driver reports as budgeted for this process.
    pub memory_budget_bytes: u64,
    /// Whether the GPU is a discrete card.
    pub discrete: bool,
    /// Apple GPU family generation rank (7 = M1 .. 10 = M4), or 0 for a non-Apple
    /// GPU. Apple silicon classifies by generation; everything else by VRAM.
    pub apple_family: u8,
}

/// The Apple GPU family generation rank a device name implies, or 0 when the name
/// is not an Apple silicon GPU. Metal reads the rank straight off the device
/// (`MTLDevice::supportsFamily`); Vulkan has no equivalent query, so a MoltenVK
/// build recovers it from the reported device name ("Apple M2 Max"). Without it
/// Apple silicon falls through `classify_tier`'s integrated branch and the two
/// backends disagree on the same GPU. `M<n>` maps to `n + 6`, matching Metal's
/// `MTLGPUFamily::Apple7` = M1.
pub fn apple_family_from_device_name(name: &str) -> u8 {
    let Some(rest) = name.strip_prefix("Apple M") else {
        return 0;
    };
    let digits: String = rest.chars().take_while(char::is_ascii_digit).collect();
    match digits.parse::<u8>() {
        Ok(n) if n >= 1 => n.saturating_add(6),
        _ => 0,
    }
}

/// Map the gathered GPU signals to a coarse performance tier. Apple silicon is
/// classified by GPU family generation (family alone cannot separate base from
/// Pro / Max / Ultra within a generation -- a working-set refinement can split
/// them later); a non-Apple integrated / low-power GPU is the lowest tier; a
/// discrete GPU is bucketed by dedicated VRAM. An unreporting device (no memory,
/// not discrete) stays `Unknown` so the resolver uses the conservative baseline.
pub fn classify_tier(input: &GpuClassInput) -> GpuTier {
    const GB: u64 = 1 << 30;
    // Apple silicon: classify by GPU family generation.
    if input.vendor == GpuVendor::Apple && input.apple_family >= 7 {
        return match input.apple_family {
            7 => GpuTier::EntryDiscrete, // M1 class
            8 => GpuTier::MidDiscrete,   // M2 class
            _ => GpuTier::HighDiscrete,  // M3 / M4 and newer
        };
    }
    // Any non-Apple integrated / low-power GPU is the lowest tier (Apple silicon
    // is unified too, but it returned above via its family branch).
    if !input.discrete {
        return GpuTier::Integrated;
    }
    // Discrete GPU: bucket by dedicated VRAM.
    match input.memory_budget_bytes {
        0 => GpuTier::Unknown,
        b if b >= 12 * GB => GpuTier::HighDiscrete,
        b if b >= 6 * GB => GpuTier::MidDiscrete,
        _ => GpuTier::EntryDiscrete,
    }
}

/// The set of operations GraphicsSystem performs on a graphics backend.
/// Implementations are thin forwarders to the inherent methods on
/// MtlContext / DxContext / VkContext.
///
/// The asset hot-reload mutators below (`update_color_lut`,
/// `rebuild_*_geometry`, `clone_static_draw_object`, etc.) are provided
/// methods that default to a no-op, so a backend implements only the reload
/// paths it actually supports.
pub trait RenderBackend: SceneControl + Send {
    /// Window / input lifecycle.
    fn window_closed(&mut self) -> bool;
    /// Confine the cursor to the window.
    fn capture_cursor(&mut self);
    /// Take the input sampled since the last call.
    fn take_input(&mut self) -> RenderInput;
    /// Block until the GPU has drained every submitted frame.
    fn wait_idle(&self);

    /// Per-frame drive. See [`FrameParams`] for the inputs.
    fn draw_frame(&mut self, params: FrameParams<'_>) -> RenderResult<()>;
    /// Push the camera's view matrix, column-major.
    fn update_view(&mut self, matrix: [[f32; 4]; 4]);

    /// Push this frame's changed model matrices, one `(draw slot, matrix)`
    /// entry per moved draw object, applied in order. Batched so the trait is
    /// crossed once per frame rather than once per entity; the caller sends
    /// only slots whose matrix actually changed. An out-of-range slot is
    /// ignored.
    fn update_models(&mut self, updates: &[(u32, [[f32; 4]; 4])]);

    /// Retire a draw object: hide it from every pass (main, shadow, velocity)
    /// and exclude it from the ray-tracing acceleration structure, so a
    /// despawned entity's slot leaves no ghost. The slot's geometry buffers are
    /// untouched; the engine's draw-slot allocator returns the index to its
    /// free list so a later `clone_static_draw_object` can recycle it. A no-op
    /// if the index is out of range.
    fn retire_draw_object(&mut self, draw_idx: usize);

    /// Skinning. `vert_bytes` and `shadow_bytes` are Metal-only payloads;
    /// DX/VK ignore them.
    fn upload_skinned(
        &mut self,
        vertices: &[SkinnedVertex],
        indices: &[u32],
        draw_objects: Vec<SkinnedDrawObject>,
        vert_bytes: &[u8],
        frag_bytes: &[u8],
        shadow_bytes: &[u8],
    ) -> RenderResult<()>;
    /// Push one skinned slot's joint matrices for this frame.
    fn update_skinned_pose(&mut self, skinned_index: usize, matrices: &[[[f32; 4]; 4]]);

    /// Attach morph-target data to the skinned draw objects, called once after
    /// `upload_skinned`: `morphs[i]` belongs to draw object `i` (instance
    /// copies share their template's data via the `Arc`). Default no-op for a
    /// backend without a morph deformation path.
    fn upload_skinned_morphs(
        &mut self,
        _morphs: Vec<Option<alloc::sync::Arc<crate::gfx::mesh_payload::PayloadMorphs>>>,
    ) {
    }

    /// Push a skinned object's current morph-target weights, sampled by the
    /// animation system each frame. A no-op when the index is out of range or
    /// the object carries no morph targets.
    fn update_morph_weights(&mut self, _skinned_index: usize, _weights: &[f32]) {}

    // Runtime skinned spawn (pre-reserved instance pool): a backend pre-reserves
    // hidden bind-pose copies at load (`SkinnedMesh.max_instances`) and reveals
    // one per skinned SpawnRequest. The default no-op implementations are a
    // fallback for a backend that has not wired runtime skinned spawn, where a
    // skinned SpawnRequest finds nothing to claim and is dropped.

    /// Reveal the pre-reserved skinned instance at `instance_index` (a hidden
    /// bind-pose copy expanded at load): show it at `model` and reset its
    /// palette to bind so it does not flash a previous occupant's pose. Which
    /// instance to use is decided by the engine's instance pool; the backend
    /// only applies it. A no-op if the index is out of range.
    fn reveal_skinned_instance(&mut self, _instance_index: usize, _model: [[f32; 4]; 4]) {}

    /// Hide a live skinned instance. The engine's instance pool returns the
    /// slot for reuse; the backend only hides it. A no-op if the index is out
    /// of range.
    fn retire_skinned_draw_object(&mut self, _skinned_index: usize) {}

    /// Push this frame's changed skinned model-to-world matrices, one
    /// `(skinned index, matrix)` entry per moved instance, applied in order
    /// (a skinned object animates in place unless something moves it). Cheap:
    /// the per-frame cull rebuild reads the object's model directly, so this
    /// just writes the fields. Out-of-range indices are ignored; default
    /// no-op for a backend without movable skinned instances.
    fn update_skinned_models(&mut self, _updates: &[(u32, [[f32; 4]; 4])]) {}

    /// Texture streaming. Albedo and normal maps share one handle-indexed pool,
    /// so every streamed texture (whatever its role) flows through these. The
    /// image carries its GPU format and mip chain: RGBA8 regenerates mips on
    /// upload, block-compressed formats upload their chain verbatim.
    fn evict_texture_slot(&mut self, slot: usize) -> Result<(), String>;
    /// Replace a texture slot's image after a streaming upload.
    fn update_texture_slot(
        &mut self,
        slot: usize,
        image: &crate::build::texture::TextureImage,
    ) -> RenderResult<()>;

    /// Mesh streaming.
    fn evict_mesh(&mut self, draw_idx: usize, retire_frame: u64) -> Result<(), String>;
    /// Upload a streamed mesh's geometry into a draw slot.
    fn upload_mesh(
        &mut self,
        draw_idx: usize,
        verts: &[Vertex],
        idxs: &[u16],
        frame: u64,
    ) -> RenderResult<()>;

    /// Seed the streamed-mesh sub-allocators with one reserved headroom block
    /// (byte ranges in the shared vertex / index buffers) instead of the
    /// per-mesh build-time regions. Used by the shrinkable-seed path: the
    /// streamed geometry is no longer baked into the buffers at build time, so
    /// the renderer hands the allocators one contiguous block sized to the
    /// cap-many resident meshes rather than the whole streamed set. Implemented
    /// on Metal + DirectX + Vulkan. Default no-op: a backend without the
    /// shrinkable seed keeps freeing each mesh's build-time region in
    /// `setup_mesh_streaming`.
    fn seed_mesh_streaming(
        &mut self,
        vtx_offset: u64,
        vtx_bytes: u64,
        idx_offset: u64,
        idx_bytes: u64,
    ) {
        let _ = (vtx_offset, vtx_bytes, idx_offset, idx_bytes);
    }

    /// Voxel-world chunk streaming. `texture_slot` and `normal_map_slot`
    /// are ignored by Metal (it binds chunk textures per draw).
    fn setup_chunk_streaming(
        &mut self,
        chunk_vtx_bytes: usize,
        chunk_idx_bytes: usize,
        texture_slot: usize,
        normal_map_slot: usize,
    ) -> RenderResult<()>;
    /// The destination draw slot comes from the engine's allocator, like
    /// `clone_static_draw_object`; the freed slot is likewise returned to it by
    /// the caller of `remove_chunk_mesh`.
    fn add_chunk_mesh(
        &mut self,
        mesh: ChunkMesh<'_>,
        dst: crate::render::draw_slot::SlotAlloc,
    ) -> RenderResult<()>;
    /// Free a streamed chunk's geometry, retiring it after `retire_frame`.
    fn remove_chunk_mesh(&mut self, draw_idx: usize, retire_frame: u64) -> Result<(), String>;
    /// Move a streamed chunk by replacing its placement matrix.
    fn set_chunk_model(&mut self, draw_idx: usize, model: [[f32; 4]; 4]) -> Result<(), String>;

    /// Device capability flags, queried from the GPU once the backend is built.
    /// Read by GraphicsSystem to gray out + disable settings rows the device
    /// cannot honor. Default: all capable, so a backend that does not report
    /// capabilities keeps every toggle live (the feature then no-ops with a
    /// warning on an incapable device, as before).
    fn capabilities(&self) -> DeviceCapabilities {
        DeviceCapabilities::ALL
    }

    /// Coarse GPU performance profile, queried once the backend is built. Read at
    /// init to pick default graphics quality on first launch. Default: `UNKNOWN`
    /// (the conservative tier), so a backend that does not report a profile never
    /// makes the resolver auto-select a high preset.
    fn gpu_profile(&self) -> GpuProfile {
        GpuProfile::UNKNOWN
    }

    /// The overlay coordinate space: the window's content size in logical,
    /// DPI-independent units (points on macOS, client pixels on Windows, window
    /// coordinates on Linux). Every backend reports the cursor in these same
    /// units, so UI hit-testing, text layout, and the overlay shader's divide to
    /// NDC all share one space regardless of the backing scale. A backend
    /// converts to attachment pixels only where a pixel rect is unavoidable,
    /// through `fullscreen::clip_rect_to_scissor`.
    ///
    /// Default `(0.0, 0.0)` for a headless backend with no window.
    fn logical_size(&self) -> (f32, f32) {
        (0.0, 0.0)
    }
    /// Per-frame draw-call / object counters. Default no-op so a backend that
    /// tracks none still satisfies the trait; all three shipping backends
    /// override it.
    fn render_stats(&self) -> RenderStats {
        RenderStats::default()
    }

    /// Show or hide the OS cursor for an in-engine UI cursor (e.g. a MainMenu),
    /// independent of camera capture. Edge-triggered by the backend, so calling
    /// it every frame with the same value is cheap. Default no-op: a backend
    /// without a free-mode cursor hide leaves the system cursor visible (DX /
    /// Vulkan today).
    fn set_ui_cursor_hidden(&mut self, hidden: bool) {
        let _ = hidden;
    }

    /// Whether the real cursor has left the window, so an in-engine UI cursor
    /// should stop drawing (windowed / borderless). The backend confines the
    /// cursor to the active screen while in fullscreen, so it reports `false`
    /// there. Default `false` (inside): backends without window-bounds tracking
    /// (DX / Vulkan today) always draw the in-engine cursor.
    fn cursor_outside_window(&self) -> bool {
        false
    }

    /// Tell the backend a togglable menu (a Screen toggled by an Escape KeyBinding)
    /// coexists with a captured camera. In this mode Escape routes to the ECS
    /// (so the menu shows/hides) instead of releasing the cursor inline, and a
    /// click never recaptures the cursor (it fires a UI action). Set once at
    /// setup. Default no-op: backends without dynamic capture (DX / Vulkan today)
    /// keep the static behavior.
    fn set_menu_mode(&mut self, on: bool) {
        let _ = on;
    }

    /// Drive cursor capture from the menu state each frame: capture for camera
    /// control, release while a menu is open. Edge-triggered by the backend.
    /// Default no-op (DX / Vulkan): they keep their startup capture decision.
    fn set_camera_capture(&mut self, capture: bool) {
        let _ = capture;
    }

    /// Supply the reflection-probe placements (from declared `ReflectionProbe`
    /// assets, or empty to auto-seed from the scene bounds). The backend bakes a
    /// cube per placement and samples the nearest for the specular reflection.
    /// Pushed once after construction. Default no-op: backends without probe
    /// support (DX / Vulkan today) keep the sky reflection.
    fn set_reflection_probes(
        &mut self,
        probes: &[crate::render::reflection_probe::ProbePlacement],
    ) {
        let _ = probes;
    }

    /// Turn display sync (vsync) on or off at runtime, applied to presentation.
    /// Edge-triggered by the backend, so calling it with the unchanged value is
    /// cheap. Default no-op: a backend that only honors vsync at init ignores
    /// runtime changes.
    fn set_vsync(&mut self, on: bool) {
        let _ = on;
    }

    /// Switch the window between windowed / borderless / fullscreen at runtime.
    /// The change flows through the backend's normal resize path (no GPU rebuild
    /// beyond the resize it triggers). Default no-op for backends without a
    /// window (embedded / preview) or that don't yet implement it.
    fn set_window_mode(&mut self, mode: crate::components::WindowMode) {
        let _ = mode;
    }

    /// Resize the window's content area at runtime (meaningful in windowed mode).
    /// Drives the same resize path as a user-dragged resize. Default no-op for
    /// backends without a window or that don't yet implement it.
    fn set_window_size(&mut self, width: u32, height: u32) {
        let _ = (width, height);
    }

    /// The display modes (pixel resolution + refresh rate) the display this
    /// backend renders to supports, unshaped (the caller dedups + sorts).
    /// Default empty: a backend that cannot enumerate (or has no window) makes
    /// the Resolution row fall back to the static preset list.
    fn display_modes(&self) -> Vec<crate::render::display_mode::DisplayMode> {
        Vec::new()
    }

    /// The mode the display is currently running, if the backend can read it.
    /// Shown by the Resolution row when the user has never chosen a mode (the
    /// display keeps its desktop mode until one is chosen). Default `None`.
    fn current_display_mode(&self) -> Option<crate::render::display_mode::DisplayMode> {
        None
    }

    /// Select the display mode to hold while the window is in fullscreen. The
    /// backend applies it whenever the window is (or becomes) fullscreen and
    /// restores the display's original mode when the window leaves fullscreen
    /// or shuts down; outside fullscreen the choice is only remembered. Default
    /// no-op: a backend without mode switching leaves the display alone.
    fn set_display_mode(&mut self, mode: crate::render::display_mode::DisplayMode) {
        let _ = mode;
    }

    /// Replace the live post-process tunables (bloom / exposure / vignette /
    /// LUT blend / FXAA). These are pushed to the bloom + composite shaders each
    /// frame, so a change takes effect on the next draw with no allocation or
    /// pipeline rebuild. Only the authored half travels here: the composite's
    /// display-output flags belong to the display the backend negotiated with
    /// at init, so a push cannot disturb them. Default no-op: a backend that
    /// only reads the tunables at init ignores runtime changes.
    fn update_post_process(&mut self, tunables: PostProcessTunables) {
        let _ = tunables;
    }

    /// Set the live ambient (IBL) light scale. Unlike the post-process params
    /// above, `ambient_intensity` lives in the shared `LightUniforms` (uploaded
    /// each frame by the main lighting pass), so it takes its own setter rather
    /// than `update_post_process`. Default no-op: only Metal mutates it live
    /// today; DirectX / Vulkan keep the init-time value (they read it at init).
    fn set_ambient_intensity(&mut self, value: f32) {
        let _ = value;
    }

    /// Replace the live directional-light set (the sun). Unlike the local
    /// lights, which ride a per-scene storage buffer sized once at init, the
    /// directional slots are a fixed-size array in the shared `LightUniforms`,
    /// so a new set is written in place: the backend re-packs the array and
    /// re-caches whatever it derived from the first light at init (the cascade
    /// shadow direction, the fog sun). Default no-op: a backend that only reads
    /// the lights at init keeps the init-time sun.
    fn update_directional_lights(&mut self, lights: &[crate::components::DirectionalLight]) {
        let _ = lights;
    }

    /// Push the gameplay movement key map. The backend resolves each canonical
    /// `InputKey` to its native key code and decodes physical key events through the
    /// map (instead of hardcoded keys), so a settings-menu rebind takes effect on
    /// the next key event. Pushed once after the backend is built and again on
    /// each rebind. Default no-op: a backend without keymap decode keeps its
    /// built-in defaults.
    fn set_keymap(&mut self, keymap: &KeyMap) {
        let _ = keymap;
    }

    /// Apply a change to the quality-feature toggles (TAA / SSAO / SSR / RT
    /// reflections / SSGI / auto-exposure) live. Unlike the post-process params,
    /// these gate render passes whose GPU resources (pipelines, render targets,
    /// ray-tracing acceleration structures) are built once at init, so applying a
    /// change rebuilds the affected resources in place rather than flipping a
    /// uniform. Default no-op: a backend that only reads these at init ignores
    /// runtime changes (DirectX / Vulkan today), so the choice persists and takes
    /// effect at the next launch there.
    fn apply_quality_settings(&mut self, settings: QualitySettings) {
        let _ = settings;
    }

    /// Set the shadow cascade re-render cadence live. The cascade scheduler reads
    /// the policy at the start of each shadow pass, so a change takes effect on the
    /// next draw with no pipeline rebuild or allocation (unlike the shadow map
    /// resolution, which is sized once at init). Default no-op: a backend that only
    /// reads the cadence at init keeps the init-time value (DirectX / Vulkan
    /// today), so the choice persists and takes effect at the next launch there.
    fn set_shadow_update(&mut self, update: crate::components::ShadowUpdate) {
        let _ = update;
    }

    /// Set the shadow distance (world units the cascades cover, capped at the
    /// camera far plane) live. The per-frame cascade-split computation reads it
    /// each draw, so a change takes effect on the next frame with no allocation or
    /// rebuild (it sizes no GPU resource, unlike the shadow map resolution).
    /// Default no-op: a backend that only reads the distance at init keeps the
    /// init-time value (DirectX / Vulkan today), so the choice persists and takes
    /// effect at the next launch there.
    fn set_shadow_distance(&mut self, distance: u32) {
        let _ = distance;
    }

    /// Set the live shadow cascade count (1..=4). The cascade-split math + the
    /// re-render schedule read it each frame and only the first `count` cascades
    /// are projected, rendered, and sampled (the array capacity stays 4), so a
    /// change takes effect on the next frame with no resize or rebuild. Default
    /// no-op: a backend that only reads the count at init keeps the init-time
    /// value (DirectX / Vulkan today), so the choice persists and takes effect at
    /// the next launch there.
    fn set_shadow_cascades(&mut self, count: u32) {
        let _ = count;
    }

    /// Update the live scalar sub-tunables of the SSAO / SSR / SSGI / auto-exposure
    /// passes (radius, intensity, distance, EV bounds, adaptation speed). Unlike
    /// `apply_quality_settings`, this rebuilds nothing: each backend re-reads these
    /// values from its stored `*Settings` structs into a per-frame uniform every
    /// draw, so mutating them takes effect on the next frame with no pipeline /
    /// target rebuild and no TAA-history reset. Only the fields of a feature that is
    /// currently on are honoured (its settings are present); a value for an off
    /// feature is ignored here and applies when the feature next turns on. The
    /// structural sub-knobs (gather resolution, ray / step counts) are NOT live and
    /// still ride `apply_quality_settings`. Default no-op: a backend that reads
    /// these only at init keeps the init-time values (DirectX / Vulkan today), so
    /// the choice persists and takes effect at the next launch there.
    fn update_quality_params(&mut self, settings: QualitySettings) {
        let _ = settings;
    }

    /// Shared atomic flag the backend polls at frame start to trigger a
    /// shader rebuild. `Some` only under `cn debug` on backends that ship
    /// hot-reload (Metal today); `None` on production runs and on backends
    /// that have not implemented hot-reload yet. The debug server reads this
    /// to forward `reload-shaders` commands; the filesystem watcher writes
    /// it directly. Default: `None`.
    fn shader_reload_flag(&self) -> Option<alloc::sync::Arc<core::sync::atomic::AtomicBool>> {
        None
    }

    /// Replace the live colour-grading LUT with a fresh `size³` RGBA8 payload.
    /// Driven by asset hot-reload (`cn debug` only). Default no-op: backends
    /// that have not implemented the swap leave the LUT bound at whatever
    /// payload was uploaded at init.
    fn update_color_lut(&mut self, size: u32, data: &[u8]) -> Result<(), String> {
        let _ = (size, data);
        Ok(())
    }

    /// `(vertex_count, index_count)` for the static draw at `draw_idx`, or
    /// `None` when the index is out of range / the backend does not expose
    /// the field. Used by asset hot-reload to detect size-changing
    /// reloads before attempting [`Self::update_mesh_geometry`], which
    /// rejects size mismatches. Default returns `None`; backends that
    /// implement the rebuild path also override this.
    fn draw_geometry_size(&self, draw_idx: usize) -> Option<(usize, usize)> {
        let _ = draw_idx;
        None
    }

    /// Per-LOD-alternate index counts for the static draw at `draw_idx`,
    /// ordered from LOD1 upward (LOD0 is reported by
    /// [`Self::draw_geometry_size`]). Returns `None` when the index is out of
    /// range or the backend does not expose its LOD layout. Used by asset
    /// hot-reload alongside [`Self::draw_geometry_size`] to detect
    /// size-changing reloads: a `.glb` that re-exports with a different LOD
    /// breakdown queues the entry for [`Self::rebuild_static_geometry`]
    /// instead of [`Self::update_mesh_geometry`]'s in-place write.
    fn draw_lod_index_counts(&self, draw_idx: usize) -> Option<Vec<usize>> {
        let _ = draw_idx;
        None
    }

    /// Rebuild the shared static-mesh vertex + index buffers, replacing the
    /// geometry of each `DrawGeometryUpdate.draw_idx` with the new
    /// vertices / indices / LOD alternates. Draws not named in `changes`
    /// keep their current geometry, copied byte-for-byte from the live
    /// buffers. The slot's `vertex_count`, `index_count`, and
    /// `lod_alternates` index offsets are rewritten as the new buffers are
    /// laid out. Driven by asset hot-reload (`cn debug` only) when a
    /// size-changing `.glb` re-export means the existing
    /// [`Self::update_mesh_geometry`] in-place write no longer fits.
    /// `wait_idle` first; the rebuild swaps the GPU buffers wholesale.
    /// Default no-op: backends that have not implemented the rebuild
    /// return `Ok(())` and the size-changing reload is logged + skipped at
    /// the caller (the existing in-place path already errored on size
    /// mismatch).
    fn rebuild_static_geometry(&mut self, changes: Vec<DrawGeometryUpdate>) -> RenderResult<()> {
        let _ = changes;
        Ok(())
    }

    /// Replace a `SkinnedMesh` draw slot's vertex + index data in place.
    /// Driven by asset hot-reload (`cn debug` only). Reuses the slot's
    /// existing vertex region + index region in the shared skinned vertex /
    /// index buffers (created once by [`Self::upload_skinned`]), so the new
    /// geometry must match the slot's init-time vertex count + index count
    /// and the new skeleton must keep the same joint count; pipelines stay
    /// untouched, only the bytes change. `vertex_base` is the init-time
    /// vertex offset (in vertex units) into the shared buffer; indices are
    /// rebased onto it before writing. Default no-op.
    fn update_skinned_mesh_geometry(
        &mut self,
        skinned_index: usize,
        vertex_base: u32,
        verts: &[SkinnedVertex],
        idxs: &[u16],
    ) -> Result<(), String> {
        let _ = (skinned_index, vertex_base, verts, idxs);
        Ok(())
    }

    /// Rebuild the shared skinned-mesh vertex + index buffers, replacing the
    /// geometry of each `SkinnedDrawGeometryUpdate.skinned_index` with the
    /// new vertices / indices. Slots not named in `changes` keep their
    /// current geometry, copied byte-for-byte from the live buffers and
    /// re-based onto the new vertex region they land in. Returns the
    /// post-rebuild layout (one [`SkinnedSlotLayout`] per slot, in
    /// `skinned_index` order) so the caller can refresh its source-map
    /// `vertex_base` / `vertex_count` / `index_count` to point at the new
    /// regions. Driven by asset hot-reload (`cn debug` only) when a
    /// size-changing `.glb` re-export means the existing
    /// [`Self::update_skinned_mesh_geometry`] in-place write no longer fits.
    /// The backend `wait_idle`s first; the rebuild swaps the GPU buffers
    /// wholesale. The skinned pipelines, shadow + velocity + SSAO + SSR
    /// variants, and `skinned_draw_objects` slot metadata
    /// (`texture_slot` / `normal_map_slot` / `material` / `joint_count`)
    /// all stay untouched; only the `index_offset` / `index_count` on each
    /// `SkinnedDrawObject` (and the buffers themselves) move. Default no-op
    /// (returns an empty layout vec): backends that have not implemented
    /// the rebuild leave the size-changing reload as logged + skipped at
    /// the caller, the same behaviour as before, since the in-place path
    /// already errored on size mismatch.
    fn rebuild_skinned_geometry(
        &mut self,
        changes: Vec<SkinnedDrawGeometryUpdate>,
    ) -> Result<Vec<SkinnedSlotLayout>, String> {
        let _ = changes;
        Ok(Vec::new())
    }

    /// Update a skinned slot's joint count and resize the backend's per-slot
    /// joint-matrix buffers to match. Driven by asset hot-reload (`cn debug`
    /// only) when a re-imported `.glb`'s skeleton has a different joint
    /// count than the slot was initialised with. Shrinking truncates the
    /// per-slot Vec; growing seeds the new entries to identity so the slot
    /// renders undeformed on the next `update_skinned_pose`. The skinned
    /// shaders consume the joints buffer through a pointer (not a fixed-
    /// size array) and use vertex-attribute-encoded joint indices, so no
    /// pipeline or shader rebuild is required for a joint-count change;
    /// only the CPU-side per-slot buffer and `SkinnedDrawObject.joint_count`
    /// change. Default no-op: backends that have not implemented the resize
    /// leave the skeleton-shape change logged + skipped at the caller.
    fn update_skinned_skeleton(
        &mut self,
        skinned_index: usize,
        new_joint_count: usize,
    ) -> Result<(), String> {
        let _ = (skinned_index, new_joint_count);
        Ok(())
    }

    /// Replace a `Mesh` draw slot's vertex + index data in place. Driven by
    /// asset hot-reload (`cn debug` only). Reuses the slot's existing offset
    /// in the shared vertex / index buffers, so the new geometry must match
    /// the slot's init-time vertex count + index count; a size-changing
    /// reload returns an error so the caller can queue
    /// [`Self::rebuild_static_geometry`] instead, which repacks the shared
    /// buffers. Each entry in
    /// `lod_alternates` (`(switch_distance, mesh-relative indices)`) is
    /// written to the matching slot's pre-allocated LOD index region; the
    /// number of LODs and each LOD's index count must match the slot's
    /// init-time layout, otherwise the call returns an error so the caller
    /// can queue [`Self::rebuild_static_geometry`]. `switch_distance` is
    /// re-stored per LOD so a JSON-side tweak to `lod_distances` propagates
    /// without a process restart. Default no-op.
    fn update_mesh_geometry(
        &mut self,
        draw_idx: usize,
        verts: &[Vertex],
        idxs: &[u16],
        lod_alternates: &[(f32, Vec<u16>)],
    ) -> Result<(), String> {
        let _ = (draw_idx, verts, idxs, lod_alternates);
        Ok(())
    }

    /// Replace the live IBL environment map with a freshly precomputed payload.
    /// `payload` is the serialised byte format emitted by
    /// `crate::build::environment_map::compile_environment_map_payload`
    /// (header + irradiance cube + prefilter mip chain), so init and hot-reload
    /// share a single byte format. Driven by asset hot-reload (`cn debug`
    /// only). Default no-op: backends that have not implemented the swap leave
    /// the IBL cubes bound at whatever payload was uploaded at init.
    fn update_environment_map(&mut self, payload: &[u8]) -> RenderResult<()> {
        let _ = payload;
        Ok(())
    }

    /// Replace the live volumetric-fog settings, or disable the fog pass when
    /// `None`. Driven by world.jsonl hot-reload (`cn debug` only). Default
    /// no-op: backends that have not implemented the swap leave the fog pass
    /// at whatever settings were resolved at init.
    ///
    /// A backend that built its fog pipeline lazily based on the world's
    /// init-time `VolumetricFog` cannot enable the pass via this call when
    /// the world started with no fog declared; re-enabling fog on a world
    /// that did not declare it at startup requires a relaunch.
    fn update_fog_settings(&mut self, settings: Option<FogSettings>) {
        let _ = settings;
    }

    /// Capture the last presented frame to a PNG at `path` and return the saved
    /// path. Driven by the `cn debug` WS `screenshot` command for headless
    /// on-GPU render verification. Default `Err`: a backend without a capture
    /// path reports it unsupported (all current backends override this).
    fn screenshot(&mut self, path: &str) -> Result<String, String> {
        let _ = path;
        Err("screenshot capture not supported on this backend".to_string())
    }

    /// Instantiate a runtime copy of an existing draw object at a new transform:
    /// re-use the source slot's geometry region (`vertex_offset` / `vertex_count`
    /// / `index_offset` / `index_count` / `base_vertex` / `lod_alternates`) and
    /// copy its texture slots, material, and cull distance, swapping only the
    /// model matrix. The new slot reuses one freed by `retire_draw_object` before
    /// growing the draw-object vec. The destination slot comes from the
    /// engine's draw-slot allocator: `Reuse` overwrites a vacated entry,
    /// `Append` grows the vec (the index always equals the current length,
    /// which implementations debug-assert). Driven by runtime entity spawn
    /// (`SpawnRequest`). The copy is non-cullable (sentinel AABB) and drawn
    /// every frame, since the init-time BVH cannot refit to admit a slot added
    /// at runtime; moving copies (the common case) opt out of the static BVH
    /// exactly like streamed chunks and held items. Default no-op (returns
    /// `Err`): backends without an implementation leave the spawn path
    /// logged + skipped at the caller.
    fn clone_static_draw_object(
        &mut self,
        src_draw_idx: usize,
        model: [[f32; 4]; 4],
        dst: crate::render::draw_slot::SlotAlloc,
    ) -> Result<(), String> {
        let _ = (src_draw_idx, model, dst);
        Err("clone_static_draw_object: not implemented on this backend".to_string())
    }

    /// Rewrite a draw slot's material parameters + texture/normal-map pool
    /// indices in place. Driven by the editor's live draw seam when a Prop edits
    /// its `material` arg. Default no-op; a backend that implements it reports
    /// [`DeviceCapabilities::rewrites_draws`], which is what the caller gates on
    /// rather than pushing an edit that would not land.
    fn set_draw_material(
        &mut self,
        draw_idx: usize,
        material: MaterialUniforms,
        texture_slot: usize,
        normal_map_slot: usize,
    ) {
        let _ = (draw_idx, material, texture_slot, normal_map_slot);
    }

    /// Rewrite a draw slot's `cull_distance` in place. Driven by the editor's
    /// live draw seam when a Prop edits its `cull_distance` arg. Default no-op,
    /// gated by the same [`DeviceCapabilities::rewrites_draws`] flag.
    fn set_draw_cull_distance(&mut self, draw_idx: usize, cull_distance: f32) {
        let _ = (draw_idx, cull_distance);
    }

    /// Append a projected-decal record at runtime, returning a stable slot
    /// index the caller hands to [`Self::remove_decal`] later. Lets a
    /// gameplay system stamp bullet holes, footprints, or other ad-hoc
    /// decals after the world has built. Backends that have not implemented
    /// the runtime path return `Err`; the caller logs and drops the request.
    fn add_decal(&mut self, record: crate::render::decal::DecalRecord) -> Result<usize, String> {
        let _ = record;
        Err("add_decal: not implemented on this backend".to_string())
    }

    /// Tombstone a runtime decal slot. The id returned by
    /// [`Self::add_decal`] becomes invalid; the next add may reuse it.
    /// Default no-op-with-Err: backends without a runtime path leave the
    /// remove logged + skipped at the caller.
    fn remove_decal(&mut self, decal_id: usize) -> Result<(), String> {
        let _ = decal_id;
        Err("remove_decal: not implemented on this backend".to_string())
    }

    /// Append a particle-emitter record at runtime, returning a stable slot
    /// index. The backend allocates the per-emitter GPU pool + atomic
    /// spawn counter (matching the init-time path) so the compute kernel
    /// can begin ticking on the next frame. Default no-op-with-Err.
    fn add_emitter(
        &mut self,
        record: crate::render::particles::ParticleEmitterRecord,
    ) -> Result<usize, String> {
        let _ = record;
        Err("add_emitter: not implemented on this backend".to_string())
    }

    /// Tombstone a runtime emitter slot and release its GPU pool +
    /// counter buffers (the GPU keeps them alive via its own refcount
    /// until any in-flight command buffer that referenced them completes).
    /// Default no-op-with-Err.
    fn remove_emitter(&mut self, emitter_id: usize) -> Result<(), String> {
        let _ = emitter_id;
        Err("remove_emitter: not implemented on this backend".to_string())
    }

    /// Rebuild the live main / instanced / shadow render pipelines from
    /// freshly compiled world-loaded shader stage bytes. Driven by asset
    /// hot-reload (`cn debug` only) when one of the captured `Shader`
    /// source files is saved or a debug-WS `reload-assets` command fires.
    /// Each `Some(bytes)` replaces the matching live pipeline (and any
    /// dependent state: bindless-texture argument encoder, cull pipeline,
    /// instanced variant, shadow variant); `None` leaves the pipeline
    /// untouched (e.g. a world without an instanced shader passes `None`
    /// for the instanced slot). The backend should build every replacement
    /// into a temporary first and only swap when every build succeeds;
    /// mirrors the safety pattern in the Metal backend's `hot_reload` so a
    /// compile error never overwrites a live pipeline with a half-built
    /// replacement. Default no-op (returns `Err`): backends without an
    /// implementation leave the world-loaded shader reload logged + skipped
    /// at the caller.
    ///
    /// Skinned-mesh variants are out of scope here: their pipelines depend
    /// on the world's `SkinnedMesh`-injected library bytes that
    /// [`Self::upload_skinned`] consumes and drops.
    fn update_world_shader_pipelines(
        &mut self,
        vert_bytes: Option<&[u8]>,
        frag_bytes: Option<&[u8]>,
        shadow_bytes: Option<&[u8]>,
        vert_instanced_bytes: Option<&[u8]>,
    ) -> Result<(), String> {
        let _ = (vert_bytes, frag_bytes, shadow_bytes, vert_instanced_bytes);
        Err("update_world_shader_pipelines: not implemented on this backend".to_string())
    }

    /// Build the render pipeline for one shader bucket from its compiled stage
    /// bytes, making draws that carry that bucket renderable. Called by the
    /// streaming pump when a scene that exclusively owns the bucket's `Shader`
    /// pins: init skipped the build, so this is where the cost lands (behind
    /// the loading screen, since the bucket counts as scene-resident content).
    /// Bucket 0 is the world default program and is never installed this way.
    ///
    /// Default no-op-with-Ok: a backend that renders every draw with the world
    /// default program has no per-bucket pipeline to build, and the bucket is
    /// resident as far as scene loading is concerned.
    fn install_world_shader(&mut self, bucket: u32, shader: ShaderBytes<'_>) -> RenderResult<()> {
        let _ = (bucket, shader);
        Ok(())
    }

    /// Release one shader bucket's render pipeline, undoing
    /// [`Self::install_world_shader`]. Called when the owning scene unpins;
    /// draws carrying the bucket stop rendering until it is installed again.
    /// Default no-op, for the same reason as above.
    fn evict_world_shader(&mut self, bucket: u32) {
        let _ = bucket;
    }

    /// The swapchain-level configuration this live backend can hot-swap a world
    /// onto, or `None` when the backend cannot reload a world in place (it must
    /// be fully rebuilt instead). Read by GraphicsSystem when a transplanted
    /// backend is handed a new world (the `cn editor` live SAVE): the swap reuses
    /// the backend via [`Self::reload_world`] only when this equals the new
    /// world's `BackendInit::swapchain_config`; a `None` or a mismatch routes to a
    /// full rebuild (recreating the window). Default `None`: DirectX / Vulkan
    /// (and any backend without a real `reload_world`) always rebuild.
    fn hot_swap_config(&self) -> Option<SwapchainConfig> {
        None
    }

    /// Re-upload a new world's GPU content onto this already-constructed backend,
    /// reusing the live device + window + swapchain instead of building a new one.
    /// Driven by the `cn editor` live SAVE: after a structural edit recompiles the
    /// blobs, GraphicsSystem transplants the running backend into the rebuilt
    /// world and calls this so the edit applies without recreating the OS window
    /// or re-initialising the GPU device. The backend waits for the GPU to idle,
    /// drops the old world's content resources, and rebuilds them from `init` on
    /// the retained hardware. Only ever called when [`Self::hot_swap_config`]
    /// reported a config matching `init.swapchain_config()`, so the swapchain
    /// (pixel format / frames-in-flight / EDR) is guaranteed unchanged. Default
    /// `Err`/unsupported: DirectX / Vulkan fall back to a full rebuild (no
    /// regression; a real implementation is Windows-pending like the rest).
    fn reload_world(&mut self, init: BackendInit<'_>) -> RenderResult<()> {
        let _ = init;
        Err(RenderError::Other(
            "reload_world: not supported on this backend".to_string(),
        ))
    }
}

// A do-nothing backend used to exercise the trait's provided (default) method
// bodies without a GPU: the smallest valid bodies for the required methods,
// no defaults overridden. Shared by this module's tests and the ops tests.
#[cfg(test)]
pub(crate) mod test_stub {
    use super::*;

    pub(crate) struct StubBackend;

    impl SceneControl for StubBackend {
        fn update_visibility(&mut self, _draw_idx: usize, _visible: bool) {}
        fn set_fade(&mut self, _fade: f32) {}
    }

    impl RenderBackend for StubBackend {
        fn window_closed(&mut self) -> bool {
            false
        }
        fn capture_cursor(&mut self) {}
        fn take_input(&mut self) -> RenderInput {
            RenderInput::default()
        }
        fn wait_idle(&self) {}
        fn draw_frame(&mut self, _params: FrameParams<'_>) -> RenderResult<()> {
            Ok(())
        }
        fn update_view(&mut self, _matrix: [[f32; 4]; 4]) {}
        fn update_models(&mut self, _updates: &[(u32, [[f32; 4]; 4])]) {}
        fn retire_draw_object(&mut self, _draw_idx: usize) {}
        fn upload_skinned(
            &mut self,
            _vertices: &[SkinnedVertex],
            _indices: &[u32],
            _draw_objects: Vec<SkinnedDrawObject>,
            _vert_bytes: &[u8],
            _frag_bytes: &[u8],
            _shadow_bytes: &[u8],
        ) -> RenderResult<()> {
            Ok(())
        }
        fn update_skinned_pose(&mut self, _skinned_index: usize, _matrices: &[[[f32; 4]; 4]]) {}
        fn evict_texture_slot(&mut self, _slot: usize) -> Result<(), String> {
            Ok(())
        }
        fn update_texture_slot(
            &mut self,
            _slot: usize,
            _image: &crate::build::texture::TextureImage,
        ) -> RenderResult<()> {
            Ok(())
        }
        fn evict_mesh(&mut self, _draw_idx: usize, _retire_frame: u64) -> Result<(), String> {
            Ok(())
        }
        fn upload_mesh(
            &mut self,
            _draw_idx: usize,
            _verts: &[Vertex],
            _idxs: &[u16],
            _frame: u64,
        ) -> RenderResult<()> {
            Ok(())
        }
        fn setup_chunk_streaming(
            &mut self,
            _chunk_vtx_bytes: usize,
            _chunk_idx_bytes: usize,
            _texture_slot: usize,
            _normal_map_slot: usize,
        ) -> RenderResult<()> {
            Ok(())
        }
        fn add_chunk_mesh(
            &mut self,
            _mesh: ChunkMesh<'_>,
            _dst: crate::render::draw_slot::SlotAlloc,
        ) -> RenderResult<()> {
            Ok(())
        }
        fn remove_chunk_mesh(
            &mut self,
            _draw_idx: usize,
            _retire_frame: u64,
        ) -> Result<(), String> {
            Ok(())
        }
        fn set_chunk_model(
            &mut self,
            _draw_idx: usize,
            _model: [[f32; 4]; 4],
        ) -> Result<(), String> {
            Ok(())
        }
    }
}

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

    use alloc::vec;
    const GB: u64 = 1 << 30;

    fn input(
        vendor: GpuVendor,
        memory_budget_bytes: u64,
        discrete: bool,
        apple_family: u8,
    ) -> GpuClassInput {
        GpuClassInput {
            vendor,
            memory_budget_bytes,
            discrete,
            apple_family,
        }
    }

    #[test]
    fn unknown_profile_is_the_conservative_default() {
        // The opposite default from capabilities: quality auto-config fails safe.
        let p = GpuProfile::default();
        assert_eq!(p.tier, GpuTier::Unknown);
        assert_eq!(p.vendor, GpuVendor::Other);
        assert_eq!(p.memory_budget_bytes, 0);
        // Unknown sorts below every real tier, so a `>=` resolver treats it as
        // the floor.
        assert!(GpuTier::Unknown < GpuTier::Integrated);
        assert!(GpuTier::Integrated < GpuTier::EntryDiscrete);
        assert!(GpuTier::EntryDiscrete < GpuTier::MidDiscrete);
        assert!(GpuTier::MidDiscrete < GpuTier::HighDiscrete);
    }

    #[test]
    fn apple_family_reads_the_generation_out_of_the_device_name() {
        // The names MoltenVK reports, mapped onto Metal's family ranks.
        assert_eq!(apple_family_from_device_name("Apple M1"), 7);
        assert_eq!(apple_family_from_device_name("Apple M2 Max"), 8);
        assert_eq!(apple_family_from_device_name("Apple M3 Pro"), 9);
        assert_eq!(apple_family_from_device_name("Apple M4 Ultra"), 10);
        // A generation past what Metal's SDK names yet still ranks above M3, so
        // a newer Mac is not demoted.
        assert!(apple_family_from_device_name("Apple M9") > 9);
    }

    #[test]
    fn non_apple_device_names_report_no_family() {
        for name in [
            "NVIDIA GeForce RTX 4090",
            "AMD Radeon RX 7900 XTX",
            "Intel(R) Arc(tm) A770",
            // Apple's own non-M naming, and a truncated / malformed report.
            "Apple A17 Pro",
            "Apple M",
            "Apple MX",
            "",
        ] {
            assert_eq!(apple_family_from_device_name(name), 0, "{name}");
        }
    }

    #[test]
    fn apple_family_from_a_name_reaches_the_same_tier_metal_does() {
        // The whole point of the name probe: a MoltenVK build must land on the
        // tier the Metal backend reports for the same silicon, not on the
        // integrated floor a zero family falls through to.
        let family = apple_family_from_device_name("Apple M2 Max");
        assert_eq!(
            classify_tier(&input(GpuVendor::Apple, 32 * GB, false, family)),
            GpuTier::MidDiscrete
        );
        assert_eq!(
            classify_tier(&input(GpuVendor::Apple, 32 * GB, false, 0)),
            GpuTier::Integrated
        );
    }

    #[test]
    fn apple_silicon_classifies_by_generation() {
        // Unified memory is large on Apple silicon, but the family generation
        // (not the working-set) decides the tier, so the huge shared budget does
        // not read as a high-VRAM discrete card.
        assert_eq!(
            classify_tier(&input(GpuVendor::Apple, 16 * GB, false, 7)),
            GpuTier::EntryDiscrete // M1
        );
        assert_eq!(
            classify_tier(&input(GpuVendor::Apple, 24 * GB, false, 8)),
            GpuTier::MidDiscrete // M2
        );
        assert_eq!(
            classify_tier(&input(GpuVendor::Apple, 48 * GB, false, 9)),
            GpuTier::HighDiscrete // M3
        );
        assert_eq!(
            classify_tier(&input(GpuVendor::Apple, 64 * GB, false, 10)),
            GpuTier::HighDiscrete // M4 and newer cap at high
        );
    }

    #[test]
    fn discrete_gpu_classifies_by_vram() {
        // An Intel-Mac AMD dGPU or a PC discrete card: vendor is not Apple and
        // there is no Apple family, so VRAM buckets the tier.
        assert_eq!(
            classify_tier(&input(GpuVendor::Nvidia, 24 * GB, true, 0)),
            GpuTier::HighDiscrete
        );
        assert_eq!(
            classify_tier(&input(GpuVendor::Amd, 8 * GB, true, 0)),
            GpuTier::MidDiscrete
        );
        assert_eq!(
            classify_tier(&input(GpuVendor::Nvidia, 4 * GB, true, 0)),
            GpuTier::EntryDiscrete
        );
        // A discrete card that reports no memory budget is left Unknown rather
        // than guessed high.
        assert_eq!(
            classify_tier(&input(GpuVendor::Amd, 0, true, 0)),
            GpuTier::Unknown
        );
    }

    #[test]
    fn integrated_gpu_is_the_lowest_tier() {
        // Non-Apple integrated part: no dedicated memory, not unified, no Apple
        // family.
        assert_eq!(
            classify_tier(&input(GpuVendor::Intel, 0, false, 0)),
            GpuTier::Integrated
        );
    }

    #[test]
    fn vram_bucket_boundaries() {
        // Boundaries are inclusive lower bounds (>= 12 GB high, >= 6 GB mid).
        assert_eq!(
            classify_tier(&input(GpuVendor::Nvidia, 12 * GB, true, 0)),
            GpuTier::HighDiscrete
        );
        assert_eq!(
            classify_tier(&input(GpuVendor::Nvidia, 12 * GB - 1, true, 0)),
            GpuTier::MidDiscrete
        );
        assert_eq!(
            classify_tier(&input(GpuVendor::Nvidia, 6 * GB, true, 0)),
            GpuTier::MidDiscrete
        );
        assert_eq!(
            classify_tier(&input(GpuVendor::Nvidia, 6 * GB - 1, true, 0)),
            GpuTier::EntryDiscrete
        );
    }

    pub(crate) use super::test_stub::StubBackend;

    const IDENTITY: [[f32; 4]; 4] = [
        [1.0, 0.0, 0.0, 0.0],
        [0.0, 1.0, 0.0, 0.0],
        [0.0, 0.0, 1.0, 0.0],
        [0.0, 0.0, 0.0, 1.0],
    ];

    // Minimal QualitySettings with every feature off, so no *Settings sub-type
    // needs constructing.
    fn stub_quality() -> QualitySettings {
        QualitySettings {
            taa: false,
            ssao: None,
            ssr: None,
            rt_reflections: None,
            ssgi: None,
            reflection_blur_scale: 1,
            auto_exposure: None,
            auto_exposure_bias_ev: 0.0,
        }
    }

    #[test]
    fn default_query_methods_report_conservative_values() {
        let backend = StubBackend;
        // Capabilities fail open: a backend that does not report keeps every
        // toggle live.
        assert!(backend.capabilities().ray_tracing);
        // Quality auto-config fails safe: the unknown/conservative profile.
        assert_eq!(backend.gpu_profile().tier, GpuTier::Unknown);
        assert_eq!(backend.gpu_profile().vendor, GpuVendor::Other);
        assert_eq!(backend.gpu_profile().memory_budget_bytes, 0);
        // Diagnostics a backend may leave to the default: zeroed here.
        assert_eq!(backend.logical_size(), (0.0, 0.0));
        assert_eq!(backend.render_stats(), RenderStats::default());
        // No window-bounds tracking: the in-engine cursor always draws.
        assert!(!backend.cursor_outside_window());
        // No display enumeration and no hot-reload flag wired.
        assert!(backend.display_modes().is_empty());
        assert!(backend.current_display_mode().is_none());
        assert!(backend.shader_reload_flag().is_none());
        // Not hot-swap-capable: a live world reload routes to a full rebuild.
        assert!(backend.hot_swap_config().is_none());
        // No geometry-size introspection for the reload size check.
        assert!(backend.draw_geometry_size(0).is_none());
        assert!(backend.draw_lod_index_counts(0).is_none());
    }

    #[test]
    fn default_mutators_are_noops_and_fallible_hooks_report_defaults() {
        let mut backend = StubBackend;

        // Runtime skinned-spawn fallbacks: nothing to reveal or hide.
        backend.reveal_skinned_instance(0, IDENTITY);
        backend.retire_skinned_draw_object(0);
        backend.update_skinned_models(&[(0, IDENTITY)]);

        // Streaming + cursor + capture no-ops.
        backend.seed_mesh_streaming(0, 0, 0, 0);
        backend.set_ui_cursor_hidden(true);
        backend.set_menu_mode(true);
        backend.set_camera_capture(true);
        backend.set_reflection_probes(&[]);

        // Presentation + window no-ops.
        backend.set_vsync(true);
        backend.set_window_mode(crate::components::WindowMode::Fullscreen);
        backend.set_window_size(1280, 720);
        backend.set_display_mode(crate::render::display_mode::DisplayMode {
            width: 1920,
            height: 1080,
            refresh_hz: 60,
        });

        // Live look + input tunable no-ops.
        backend.update_post_process(PostProcessTunables::DEFAULT);
        backend.set_ambient_intensity(1.0);
        backend.set_keymap(&KeyMap::default());
        backend.apply_quality_settings(stub_quality());
        backend.update_quality_params(stub_quality());
        backend.set_shadow_update(crate::components::ShadowUpdate::EveryFrame);
        backend.set_shadow_distance(200);
        backend.set_shadow_cascades(3);
        backend.update_fog_settings(None);
        backend.update_directional_lights(&[]);
        backend.set_draw_material(0, MaterialUniforms::DEFAULT, 0, 0);
        backend.set_draw_cull_distance(0, 50.0);

        // Fallible hot-reload hooks that succeed by default (no-op Ok).
        assert!(backend.update_color_lut(2, &[0u8; 32]).is_ok());
        assert!(backend.rebuild_static_geometry(vec![]).is_ok());
        assert!(backend.update_skinned_mesh_geometry(0, 0, &[], &[]).is_ok());
        assert!(backend.rebuild_skinned_geometry(vec![]).unwrap().is_empty());
        assert!(backend.update_skinned_skeleton(0, 0).is_ok());
        assert!(backend.update_mesh_geometry(0, &[], &[], &[]).is_ok());
        assert!(backend.update_environment_map(&[]).is_ok());

        // Fallible hooks a bare backend does not implement: they report Err.
        assert!(backend.screenshot("unused.png").is_err());
        assert!(
            backend
                .clone_static_draw_object(
                    0,
                    IDENTITY,
                    crate::render::draw_slot::SlotAlloc::Append(0)
                )
                .is_err()
        );
        assert!(backend.add_decal(stub_decal()).is_err());
        assert!(backend.remove_decal(0).is_err());
        assert!(backend.add_emitter(stub_emitter()).is_err());
        assert!(backend.remove_emitter(0).is_err());
        assert!(
            backend
                .update_world_shader_pipelines(None, None, None, None)
                .is_err()
        );
    }

    // A minimal empty-world BackendInit borrowing `window`, for exercising the
    // default `reload_world`. Empty slices are `'static`; the only real borrow
    // is the window args.
    fn empty_backend_init(window: &crate::components::Window) -> BackendInit<'_> {
        BackendInit::minimal(window, alloc::vec::Vec::new())
    }

    #[test]
    fn default_reload_world_is_unsupported() {
        // A backend without a real reload path reports the swap unsupported, so
        // the caller falls back to a full rebuild.
        let mut backend = StubBackend;
        let window = crate::components::Window::default();
        assert!(backend.reload_world(empty_backend_init(&window)).is_err());
    }

    fn stub_decal() -> crate::render::decal::DecalRecord {
        crate::render::decal::DecalRecord {
            model: IDENTITY,
            inv_model: IDENTITY,
            texture_slot: 0,
            tint: [1.0; 4],
        }
    }

    fn stub_emitter() -> crate::render::particles::ParticleEmitterRecord {
        crate::render::particles::ParticleEmitterRecord {
            texture_slot: 0,
            position: [0.0; 3],
            direction: [0.0, 1.0, 0.0],
            spread_cos: 1.0,
            speed_min: 0.0,
            speed_max: 1.0,
            lifetime_min: 0.0,
            lifetime_max: 1.0,
            gravity: [0.0, -9.8, 0.0],
            spawn_rate: 1.0,
            max_particles: 1,
            size_start: 1.0,
            size_end: 1.0,
            color_start: [1.0; 4],
            color_end: [1.0; 4],
        }
    }
}