pane_ui 0.1.0

use crate::input::Input;
use crate::textures::{TextureId, TextureRegistry, UvRect};
use glyphon::{
    Attrs, Buffer, Cache, Color as GlyphColor, Family, FontSystem, Metrics, Shaping, SwashCache,
    TextArea, TextAtlas, TextBounds, TextRenderer, Viewport,
};
use std::collections::HashMap;
use std::sync::Arc;
use winit::{
    application::ApplicationHandler,
    event::WindowEvent,
    event_loop::{ActiveEventLoop, EventLoop},
    window::{Window, WindowId},
};

// ── Scaling ───────────────────────────────────────────────────────────────────

/// Convert a grid-space point to normalised device coordinates `[x, y]`.
///
/// `gx`/`gy` are in the 1080-unit grid (origin at screen centre, y-up negative).
/// `pw`/`ph` are the physical pixel dimensions of the render target.
#[must_use]
pub fn to_ndc(gx: f32, gy: f32, pw: f32, ph: f32) -> [f32; 2] {
    let gw = grid_width(pw, ph);
    let unit = ph / 1080.0;
    let ox = gw.mul_add(-unit, pw) * 0.5;
    let px = gw.mul_add(0.5, gx).mul_add(unit, ox);
    let py = (gy + 540.0) * unit;
    [(px / pw).mul_add(2.0, -1.0), (py / ph).mul_add(-2.0, 1.0)]
}

/// Return the grid-space width for a window of size `pw × ph` pixels.
///
/// Height is always 1080 units; width scales proportionally with the aspect ratio.
#[must_use]
pub fn grid_width(pw: f32, ph: f32) -> f32 {
    (pw / ph) * 1080.0
}

// ── Color ─────────────────────────────────────────────────────────────────────

/// Linear RGBA colour with components in `0.0..=1.0`.
#[derive(Debug, Clone, Copy, PartialEq, serde::Deserialize)]
pub struct Color {
    pub r: f32,
    pub g: f32,
    pub b: f32,
    pub a: f32,
}

impl Color {
    /// Construct a colour from linear RGBA components.
    #[must_use]
    pub const fn rgba(r: f32, g: f32, b: f32, a: f32) -> Self {
        Self { r, g, b, a }
    }
    pub const WHITE: Self = Self::rgba(1.0, 1.0, 1.0, 1.0);

    fn to_glyph(self) -> GlyphColor {
        #[expect(
            clippy::cast_possible_truncation,
            clippy::cast_sign_loss,
            reason = "value is clamped to 0.0..=255.0 and rounded before cast"
        )]
        fn channel(v: f32) -> u8 {
            (v * 255.0).clamp(0.0, 255.0).round() as u8
        }
        GlyphColor::rgba(
            channel(self.r),
            channel(self.g),
            channel(self.b),
            channel(self.a),
        )
    }
}

// ── IDs / Rects ───────────────────────────────────────────────────────────────

/// Axis-aligned rectangle in grid coordinates.
#[derive(Debug, Clone, Copy)]
pub struct Rect {
    pub x: f32,
    pub y: f32,
    pub w: f32,
    pub h: f32,
}

impl Rect {
    /// Construct a rect from its top-left corner and dimensions (grid units).
    #[must_use]
    pub const fn new(x: f32, y: f32, w: f32, h: f32) -> Self {
        Self { x, y, w, h }
    }

    /// Return `true` if the grid-space point `(px, py)` falls inside this rect.
    #[must_use]
    pub fn contains(&self, px: f32, py: f32) -> bool {
        px >= self.x && px <= self.x + self.w && py >= self.y && py <= self.y + self.h
    }
}

/// Opaque handle to a compiled render pipeline, returned by [`Pane::register_shader`].
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub struct ShaderId(usize);

impl ShaderId {
    /// Wrap a pipeline-vec index. Only called from [`Pane::register_shader`].
    pub(crate) const fn new(index: usize) -> Self {
        Self(index)
    }
    /// Unwrap to the pipeline-vec index for use in draw calls.
    pub(crate) const fn index(self) -> usize {
        self.0
    }
}

/// Axis-aligned scissor rectangle in grid coordinates.
///
/// When a batch has a `ClipRect`, fragments outside it are discarded by the GPU
/// scissor test and by glyphon's `TextBounds`. Both use `clip_to_pixels` to
/// convert to physical pixel coordinates.
#[derive(Debug, Clone, Copy, PartialEq)]
pub struct ClipRect {
    pub x: f32,
    pub y: f32,
    pub w: f32,
    pub h: f32,
}

// ── Coordinate helpers ────────────────────────────────────────────────────────

/// Convert a [`ClipRect`] (grid coordinates) to pixel-space `(left, top, right, bottom)`.
/// This is the single source of truth used by both the text bounds and the scissor rect.
fn clip_to_pixels(c: ClipRect, pw: f32, ph: f32) -> (f32, f32, f32, f32) {
    let unit = ph / 1080.0;
    let gw = grid_width(pw, ph);
    let ox = gw.mul_add(-unit, pw) * 0.5;
    let left = gw.mul_add(0.5, c.x).mul_add(unit, ox);
    let top = (c.y + 540.0) * unit;
    let right = gw.mul_add(0.5, c.x + c.w).mul_add(unit, ox);
    let bottom = (c.y + c.h + 540.0) * unit;
    (left, top, right, bottom)
}

// ── Batch ─────────────────────────────────────────────────────────────────────

struct Batch {
    shader: ShaderId,
    verts: Vec<[f32; 2]>,
    color: Color,
    z: f32,
    clip: Option<ClipRect>,
    texture: Option<TextureId>,
    uv_rect: Option<UvRect>,
    corner_radius: f32,
    border_width: f32,
    state: [f32; 4],
}

// ── Text Alignment ────────────────────────────────────────────────────────────

/// Horizontal alignment for text drawn inside a fixed-width box.
#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Deserialize, Default)]
pub enum TextAlign {
    Left,
    #[default]
    Center,
    Right,
}

/// Per-batch draw metadata: (shader, vertex range, clip rect, texture, z-depth).
type BatchRange = (
    ShaderId,
    std::ops::Range<u32>,
    Option<ClipRect>,
    Option<TextureId>,
    f32,
);

/// Parameters for a single text draw call.
pub(crate) struct TextDraw<'a> {
    pub text: &'a str,
    pub x: f32,
    pub y: f32,
    pub w: f32,
    pub size: f32,
    pub color: Color,
    pub align: TextAlign,
    pub font: Option<&'a str>,
    pub bold: bool,
    pub italic: bool,
    pub clip: Option<ClipRect>,
    pub z: f32,
}

/// GPU frame resources passed to `Pane::render`.
pub(crate) struct GpuFrame<'a> {
    pub encoder: &'a mut wgpu::CommandEncoder,
    pub view: &'a wgpu::TextureView,
    pub device: &'a wgpu::Device,
    pub queue: &'a wgpu::Queue,
}

// ── Scene ─────────────────────────────────────────────────────────────────────

struct TextEntry {
    text: String,
    x: f32,
    y: f32,
    w: f32,
    size: f32,
    color: Color,
    align: TextAlign,
    font: Option<String>,
    bold: bool,
    italic: bool,
    clip: Option<ClipRect>,
    z: f32,
}

/// Batched draw list assembled each frame before it is handed to `Pane::render`.
///
/// Widgets append geometry via the `push_*` / `text*` methods.  The renderer
/// sorts by z, uploads all vertices in one write, then issues draw calls.
/// Call [`Scene::clear`] at the start of each frame to reuse the allocation.
pub struct Scene {
    batches: Vec<Batch>,
    texts: Vec<TextEntry>,
}

impl Scene {
    /// Create an empty scene.
    #[must_use]
    pub const fn new() -> Self {
        Self {
            batches: Vec::new(),
            texts: Vec::new(),
        }
    }

    /// Push a plain geometry batch (no border, no widget state).
    ///
    /// Shorthand for [`push_widget`](Scene::push_widget) with `border_width = 0` and `state = [0; 4]`.
    pub fn push_full(
        &mut self,
        verts: Vec<[f32; 2]>,
        shader: ShaderId,
        color: Color,
        z: f32,
        clip: Option<ClipRect>,
        corner_radius: f32,
    ) {
        self.push_widget(verts, shader, color, z, clip, corner_radius, 0.0, [0.0; 4]);
    }

    /// Push a widget geometry batch with full control over all vertex attributes.
    ///
    /// `state` is passed through to the shader as `@location(5)`:
    /// `[hovered, pressed, focused, disabled]`, each `0.0` or `1.0`.
    pub fn push_widget(
        &mut self,
        verts: Vec<[f32; 2]>,
        shader: ShaderId,
        color: Color,
        z: f32,
        clip: Option<ClipRect>,
        corner_radius: f32,
        border_width: f32,
        state: [f32; 4],
    ) {
        self.batches.push(Batch {
            shader,
            verts,
            color,
            z,
            clip,
            texture: None,
            uv_rect: None,
            corner_radius,
            border_width,
            state,
        });
    }

    /// Push a textured quad using the full texture (UV `0,0 → 1,1`).
    pub fn push_image(
        &mut self,
        rect: Rect,
        shader: ShaderId,
        texture: TextureId,
        z: f32,
        clip: Option<ClipRect>,
    ) {
        self.push_image_uv(rect, shader, texture, z, clip, None);
    }

    /// Push a textured quad with an optional sub-rect UV override (for sprite sheets / GIF frames).
    pub fn push_image_uv(
        &mut self,
        rect: Rect,
        shader: ShaderId,
        texture: TextureId,
        depth: f32,
        clip: Option<ClipRect>,
        uv_rect: Option<UvRect>,
    ) {
        let (rx, ry, rw, rh) = (rect.x, rect.y, rect.w, rect.h);
        let verts = vec![
            [rx, ry],
            [rx + rw, ry],
            [rx, ry + rh],
            [rx + rw, ry],
            [rx + rw, ry + rh],
            [rx, ry + rh],
        ];
        self.batches.push(Batch {
            shader,
            verts,
            color: Color::WHITE,
            z: depth,
            clip,
            texture: Some(texture),
            uv_rect,
            corner_radius: 0.0,
            border_width: 0.0,
            state: [0.0; 4],
        });
    }

    /// Draw center-aligned text in a box of width `w` at grid position `(x, y)`.
    pub fn text(&mut self, text: &str, x: f32, y: f32, w: f32, size: f32, color: Color) {
        self.push_text(&TextDraw {
            text,
            x,
            y,
            w,
            size,
            color,
            align: TextAlign::Center,
            font: None,
            bold: false,
            italic: false,
            clip: None,
            z: 0.5,
        });
    }

    /// Draw left-aligned text starting at grid position `(x, y)` with no width constraint.
    pub fn text_left(&mut self, text: &str, x: f32, y: f32, size: f32, color: Color) {
        self.push_text(&TextDraw {
            text,
            x,
            y,
            w: 0.0,
            size,
            color,
            align: TextAlign::Left,
            font: None,
            bold: false,
            italic: false,
            clip: None,
            z: 0.5,
        });
    }

    /// Full control: explicit alignment, optional font family, bold, italic, clip, z-order.
    pub(crate) fn push_text(&mut self, p: &TextDraw<'_>) {
        self.texts.push(TextEntry {
            text: p.text.to_string(),
            x: p.x,
            y: p.y,
            w: p.w,
            size: p.size,
            color: p.color,
            align: p.align,
            font: p.font.map(std::string::ToString::to_string),
            bold: p.bold,
            italic: p.italic,
            clip: p.clip,
            z: p.z,
        });
    }

    /// Discard all queued geometry and text — call at the start of each frame.
    pub fn clear(&mut self) {
        self.batches.clear();
        self.texts.clear();
    }
    /// Return `true` if no geometry or text has been queued this frame.
    #[must_use]
    pub const fn is_empty(&self) -> bool {
        self.batches.is_empty() && self.texts.is_empty()
    }
}

impl Default for Scene {
    fn default() -> Self {
        Self::new()
    }
}

// ── Vertex Layout ─────────────────────────────────────────────────────────────

const VERT_SIZE: u64 = 64;
const MAX_VERTS: u64 = 2_000_000;

// Globals uniform: [time, dt, frame, cursor_x, cursor_y, cursor_pressed, pad0, pad1]
const GLOBALS_SIZE: u64 = (std::mem::size_of::<f32>() * 8) as u64;

// ── Text System ───────────────────────────────────────────────────────────────

struct TextSystem {
    font_system: FontSystem,
    swash_cache: SwashCache,
    _cache: Cache,
    atlas: TextAtlas,
    renderer: TextRenderer,
    viewport: Viewport,
}

impl TextSystem {
    fn new(device: &wgpu::Device, queue: &wgpu::Queue, format: wgpu::TextureFormat) -> Self {
        let font_system = FontSystem::new();
        let swash_cache = SwashCache::new();
        let cache = Cache::new(device);
        let viewport = Viewport::new(device, &cache);
        let mut atlas = TextAtlas::new(device, queue, &cache, format);
        let renderer =
            TextRenderer::new(&mut atlas, device, wgpu::MultisampleState::default(), None);
        Self {
            font_system,
            swash_cache,
            _cache: cache,
            atlas,
            renderer,
            viewport,
        }
    }

    fn build_buffer(&mut self, t: &TextEntry, unit: f32, pw: f32, ph: f32) -> Buffer {
        let font_size = t.size * unit;
        let use_left = matches!(t.align, TextAlign::Left) || t.w <= 0.0;
        let buf_w = if use_left { pw } else { t.w * unit };
        let family = t.font.as_deref().map_or(Family::SansSerif, Family::Name);
        let weight = if t.bold {
            glyphon::Weight::BOLD
        } else {
            glyphon::Weight::NORMAL
        };
        let style = if t.italic {
            glyphon::Style::Italic
        } else {
            glyphon::Style::Normal
        };
        let mut buf = Buffer::new(
            &mut self.font_system,
            Metrics::new(font_size, font_size * 1.2),
        );
        buf.set_size(&mut self.font_system, Some(buf_w), Some(ph));
        buf.set_text(
            &mut self.font_system,
            &t.text,
            &Attrs::new().family(family).weight(weight).style(style),
            Shaping::Advanced,
            None,
        );
        let align = match t.align {
            TextAlign::Left => glyphon::cosmic_text::Align::Left,
            TextAlign::Center => glyphon::cosmic_text::Align::Center,
            TextAlign::Right => glyphon::cosmic_text::Align::Right,
        };
        for line in &mut buf.lines {
            line.set_align(Some(align));
        }
        buf.shape_until_scroll(&mut self.font_system, false);
        buf
    }

    fn build_text_area<'a>(
        t: &TextEntry,
        buf: &'a Buffer,
        unit: f32,
        pw: f32,
        ph: f32,
        gw: f32,
        ox: f32,
    ) -> TextArea<'a> {
        let bounds = t.clip.map_or_else(
            || TextBounds {
                left: 0,
                top: 0,
                right: pw.max(0.0).trunc() as i32,
                bottom: ph.max(0.0).trunc() as i32,
            },
            |c| {
                let (l, t, r, b) = clip_to_pixels(c, pw, ph);
                TextBounds {
                    left: l.trunc() as i32,
                    top: t.trunc() as i32,
                    right: r.trunc() as i32,
                    bottom: b.trunc() as i32,
                }
            },
        );
        TextArea {
            buffer: buf,
            left: gw.mul_add(0.5, t.x).mul_add(unit, ox),
            top: (t.y + 540.0) * unit,
            scale: 1.0,
            bounds,
            default_color: t.color.to_glyph(),
            custom_glyphs: &[],
        }
    }

    /// Returns the x offset **in pixels** (from the buffer left) where the cursor
    /// at `cursor_byte` falls, using the exact same buffer configuration that the
    /// renderer uses — same `buf_w`, same `ph`, same alignment — so glyph positions
    /// are guaranteed to match the rendered text.  The alignment offset is already
    /// baked into the returned value; add it to the text origin to get screen x.
    pub(crate) fn measure_cursor_px(
        &mut self,
        text: &str,
        font_size_px: f32,
        buf_w_px: f32,
        _ph: f32,
        align: TextAlign,
        cursor_byte: usize,
        bold: bool,
        italic: bool,
        font: Option<&str>,
    ) -> f32 {
        let family = font.map_or(Family::SansSerif, Family::Name);
        let weight = if bold {
            glyphon::Weight::BOLD
        } else {
            glyphon::Weight::NORMAL
        };
        let style = if italic {
            glyphon::Style::Italic
        } else {
            glyphon::Style::Normal
        };
        let mut buf = Buffer::new(
            &mut self.font_system,
            Metrics::new(font_size_px, font_size_px * 1.2),
        );
        // Use the same width as build_buffer. Height is unbounded — we only
        // care about x positions on the first line, and omitting height avoids
        // any scroll-based clipping that could shift glyph layout.
        buf.set_size(&mut self.font_system, Some(buf_w_px), None);
        buf.set_text(
            &mut self.font_system,
            text,
            &Attrs::new().family(family).weight(weight).style(style),
            Shaping::Advanced,
            None,
        );
        let cosmic_align = match align {
            TextAlign::Left => glyphon::cosmic_text::Align::Left,
            TextAlign::Center => glyphon::cosmic_text::Align::Center,
            TextAlign::Right => glyphon::cosmic_text::Align::Right,
        };
        for line in &mut buf.lines {
            line.set_align(Some(cosmic_align));
        }
        buf.shape_until_scroll(&mut self.font_system, false);

        for run in buf.layout_runs() {
            let glyphs = run.glyphs;
            for (i, glyph) in glyphs.iter().enumerate() {
                if cursor_byte >= glyph.start && cursor_byte < glyph.end {
                    return glyph.x;
                }
                if cursor_byte == glyph.end {
                    return glyphs.get(i + 1).map_or(glyph.x + glyph.w, |next| next.x);
                }
            }
        }
        // Cursor past all glyphs (empty text or end of text).
        buf.layout_runs()
            .last()
            .and_then(|r| r.glyphs.last())
            .map_or_else(
                || match align {
                    TextAlign::Center => buf_w_px * 0.5,
                    TextAlign::Right => buf_w_px,
                    TextAlign::Left => 0.0,
                },
                |g| g.x + g.w,
            )
    }

    fn render(
        &mut self,
        device: &wgpu::Device,
        queue: &wgpu::Queue,
        pass: &mut wgpu::RenderPass,
        texts: &[TextEntry],
        pw: f32,
        ph: f32,
    ) {
        if texts.is_empty() {
            return;
        }
        let unit = ph / 1080.0;
        let ox = grid_width(pw, ph).mul_add(-unit, pw) * 0.5; // horizontal letterbox offset, mirrors to_ndc
        #[expect(
            clippy::cast_possible_truncation,
            clippy::cast_sign_loss,
            reason = "viewport dimensions are clamped to >=0 before cast"
        )]
        self.viewport.update(
            queue,
            glyphon::Resolution {
                width: pw.max(0.0) as u32,
                height: ph.max(0.0) as u32,
            },
        );

        let buffers: Vec<Buffer> = texts
            .iter()
            .map(|t| self.build_buffer(t, unit, pw, ph))
            .collect();

        let gw = grid_width(pw, ph);
        let areas: Vec<TextArea> = texts
            .iter()
            .zip(buffers.iter())
            .map(|(t, buf)| Self::build_text_area(t, buf, unit, pw, ph, gw, ox))
            .collect();

        self.atlas.trim();
        if let Err(e) = self.renderer.prepare(
            device,
            queue,
            &mut self.font_system,
            &mut self.atlas,
            &self.viewport,
            areas,
            &mut self.swash_cache,
        ) {
            eprintln!("[pane] text prepare error: {e}");
            return;
        }
        if let Err(e) = self.renderer.render(&self.atlas, &self.viewport, pass) {
            eprintln!("[pane] text render error: {e}");
        }
    }
}

// ── Pane (standalone + overlay renderer) ─────────────────────────────────────
//
// Owns pipelines, vertex buffer, textures, and text system.
// In overlay mode the caller provides device/queue each frame.
// In standalone mode Pane owns device/queue/surface directly.
//
// GpuResources bundles the fields that are always initialised together and are
// always Some after construction. This eliminates Option-unwrap noise throughout
// the impl — if you have a &GpuResources the compiler knows it exists.
//
// StandaloneReady holds the four fields that only exist in standalone mode.
// Grouping them in an Option<StandaloneReady> means the compiler enforces that
// they are all-present or all-absent, removing the risk of partial init.

struct StandaloneReady {
    surface: wgpu::Surface<'static>,
    config: wgpu::SurfaceConfiguration,
    device: Arc<wgpu::Device>,
    queue: Arc<wgpu::Queue>,
}

struct GpuResources {
    vertex_buf: wgpu::Buffer,
    text: TextSystem,
    text_overlay: TextSystem,
    globals_bgl: wgpu::BindGroupLayout,
    globals_buf: wgpu::Buffer,
    globals_bg: wgpu::BindGroup,
}

/// wgpu renderer for a single render target.
///
/// Operates in two modes:
/// - **Overlay** — created with [`Pane::new`]; the caller owns the surface and
///   supplies a `GpuFrame` each call to `Pane::render`.
/// - **Standalone** — created internally by [`run`]; `Pane` owns the window,
///   surface, device, and queue and drives its own event loop.
pub struct Pane {
    pipelines: Vec<wgpu::RenderPipeline>,
    shader_cache: HashMap<String, ShaderId>,
    gpu: Option<GpuResources>, // None only in standalone before init()
    format: wgpu::TextureFormat,
    standalone: Option<StandaloneReady>, // None in overlay mode
    total_time: f32,
    frame_count: u32,
}

impl Pane {
    const fn gpu(&self) -> &GpuResources {
        self.gpu.as_ref().unwrap()
    }
    const fn gpu_mut(&mut self) -> &mut GpuResources {
        self.gpu.as_mut().unwrap()
    }

    fn make_gpu(
        device: &wgpu::Device,
        queue: &wgpu::Queue,
        format: wgpu::TextureFormat,
    ) -> GpuResources {
        let vertex_buf = device.create_buffer(&wgpu::BufferDescriptor {
            label: None,
            size: MAX_VERTS * VERT_SIZE,
            usage: wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
            mapped_at_creation: false,
        });
        let globals_bgl = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
            label: None,
            entries: &[wgpu::BindGroupLayoutEntry {
                binding: 0,
                visibility: wgpu::ShaderStages::VERTEX_FRAGMENT,
                ty: wgpu::BindingType::Buffer {
                    ty: wgpu::BufferBindingType::Uniform,
                    has_dynamic_offset: false,
                    min_binding_size: None,
                },
                count: None,
            }],
        });
        let globals_buf = device.create_buffer(&wgpu::BufferDescriptor {
            label: None,
            size: GLOBALS_SIZE,
            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
            mapped_at_creation: false,
        });
        let globals_bg = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: None,
            layout: &globals_bgl,
            entries: &[wgpu::BindGroupEntry {
                binding: 0,
                resource: globals_buf.as_entire_binding(),
            }],
        });
        GpuResources {
            vertex_buf,
            text: TextSystem::new(device, queue, format),
            text_overlay: TextSystem::new(device, queue, format),
            globals_bgl,
            globals_buf,
            globals_bg,
        }
    }

    /// Create for overlay mode — caller provides device/queue.
    #[must_use]
    pub fn new(device: &wgpu::Device, queue: &wgpu::Queue, format: wgpu::TextureFormat) -> Self {
        Self {
            pipelines: Vec::new(),
            shader_cache: HashMap::new(),
            gpu: Some(Self::make_gpu(device, queue, format)),
            format,
            standalone: None,
            total_time: 0.0,
            frame_count: 0,
        }
    }

    /// Create an uninitialised shell for standalone mode. Call `init()` after window is ready.
    fn new_standalone() -> Self {
        Self {
            pipelines: Vec::new(),
            shader_cache: HashMap::new(),
            gpu: None,
            format: wgpu::TextureFormat::Bgra8UnormSrgb,
            standalone: None,
            total_time: 0.0,
            frame_count: 0,
        }
    }

    async fn init(&mut self, window: Arc<Window>) -> Result<(), String> {
        let size = window.inner_size();
        let instance = wgpu::Instance::default();
        let surface = instance
            .create_surface(window)
            .map_err(|e| format!("[pane] failed to create surface: {e}"))?;
        let adapter = instance
            .request_adapter(&wgpu::RequestAdapterOptions {
                compatible_surface: Some(&surface),
                ..Default::default()
            })
            .await
            .map_err(|e| format!("[pane] no compatible GPU adapter found: {e}"))?;
        let (device, queue) = adapter
            .request_device(&wgpu::DeviceDescriptor::default())
            .await
            .map_err(|e| format!("[pane] failed to acquire GPU device: {e}"))?;
        let cap = surface.get_capabilities(&adapter);
        // Prefer an sRGB format so all colors (textures, clear color, shaders) are
        // gamma-corrected automatically by the GPU. Fall back to whatever is available.
        let format = cap
            .formats
            .iter()
            .find(|f| f.is_srgb())
            .copied()
            .unwrap_or(cap.formats[0]);
        let config = wgpu::SurfaceConfiguration {
            usage: wgpu::TextureUsages::RENDER_ATTACHMENT,
            format,
            width: size.width,
            height: size.height,
            present_mode: wgpu::PresentMode::Fifo,
            alpha_mode: cap.alpha_modes[0],
            view_formats: vec![],
            desired_maximum_frame_latency: 2,
        };
        surface.configure(&device, &config);

        let device = Arc::new(device);
        let queue = Arc::new(queue);
        self.gpu = Some(Self::make_gpu(&device, &queue, format));
        self.format = format;
        self.standalone = Some(StandaloneReady {
            surface,
            config,
            device,
            queue,
        });
        Ok(())
    }

    /// Advance `total_time`/`frame_count`, write the globals uniform, and tick textures.
    /// Must be called once per frame regardless of whether the scene is empty.
    fn advance_time(
        &mut self,
        dt: f32,
        pw: f32,
        ph: f32,
        cursor: [f32; 3],
        queue: &wgpu::Queue,
        tex_reg: &mut TextureRegistry,
    ) {
        self.total_time += dt;
        self.frame_count = self.frame_count.wrapping_add(1);
        let [cx, cy] = to_ndc(cursor[0], cursor[1], pw, ph);
        let globals_data: [f32; 8] = [
            self.total_time,
            dt,
            self.frame_count as f32,
            cx,
            cy,
            cursor[2],
            0.0,
            0.0,
        ];
        queue.write_buffer(
            &self.gpu().globals_buf,
            0,
            bytemuck::cast_slice(&globals_data),
        );
        tex_reg.update(dt);
    }

    /// Return the wgpu device when running in standalone mode, or `None` in overlay mode.
    pub(crate) fn device(&self) -> Option<&Arc<wgpu::Device>> {
        self.standalone.as_ref().map(|s| &s.device)
    }

    /// Return the wgpu queue when running in standalone mode, or `None` in overlay mode.
    pub(crate) fn queue(&self) -> Option<&Arc<wgpu::Queue>> {
        self.standalone.as_ref().map(|s| &s.queue)
    }

    /// Compile a WGSL shader and register it as a render pipeline, returning its [`ShaderId`].
    ///
    /// Results are cached by source string, so calling this with the same WGSL twice is free.
    /// The pipeline is configured for the vertex layout used by [`Scene`]: positions, UVs,
    /// colour, corner radius / border width, widget size, and widget state.
    pub fn register_shader(
        &mut self,
        device: &wgpu::Device,
        wgsl: &str,
        tex_reg: &TextureRegistry,
    ) -> ShaderId {
        if let Some(&id) = self.shader_cache.get(wgsl) {
            return id;
        }

        let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
            label: None,
            source: wgpu::ShaderSource::Wgsl(wgsl.into()),
        });
        let layout = device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
            label: None,
            bind_group_layouts: &[
                Some(&tex_reg.bind_group_layout),
                Some(&self.gpu().globals_bgl),
            ],
            immediate_size: 0,
        });
        let pipeline = device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
            label: None,
            layout: Some(&layout),
            vertex: wgpu::VertexState {
                module: &shader,
                entry_point: Some("vs"),
                buffers: &[wgpu::VertexBufferLayout {
                    array_stride: VERT_SIZE,
                    step_mode: wgpu::VertexStepMode::Vertex,
                    attributes: &[
                        wgpu::VertexAttribute {
                            offset: 0,
                            shader_location: 0,
                            format: wgpu::VertexFormat::Float32x2,
                        },
                        wgpu::VertexAttribute {
                            offset: 8,
                            shader_location: 1,
                            format: wgpu::VertexFormat::Float32x2,
                        },
                        wgpu::VertexAttribute {
                            offset: 16,
                            shader_location: 2,
                            format: wgpu::VertexFormat::Float32x4,
                        },
                        wgpu::VertexAttribute {
                            offset: 32,
                            shader_location: 3,
                            format: wgpu::VertexFormat::Float32x2,
                        },
                        wgpu::VertexAttribute {
                            offset: 40,
                            shader_location: 4,
                            format: wgpu::VertexFormat::Float32x2,
                        },
                        wgpu::VertexAttribute {
                            offset: 48,
                            shader_location: 5,
                            format: wgpu::VertexFormat::Float32x4,
                        },
                    ],
                }],
                compilation_options: wgpu::PipelineCompilationOptions::default(),
            },
            fragment: Some(wgpu::FragmentState {
                module: &shader,
                entry_point: Some("fs"),
                targets: &[Some(wgpu::ColorTargetState {
                    format: self.format,
                    blend: Some(wgpu::BlendState::ALPHA_BLENDING),
                    write_mask: wgpu::ColorWrites::ALL,
                })],
                compilation_options: wgpu::PipelineCompilationOptions::default(),
            }),
            primitive: wgpu::PrimitiveState::default(),
            depth_stencil: None,
            multisample: wgpu::MultisampleState::default(),
            multiview_mask: None,
            cache: None,
        });

        let id = ShaderId::new(self.pipelines.len());
        self.pipelines.push(pipeline);
        self.shader_cache.insert(wgsl.to_string(), id);
        id
    }

    /// Returns the cursor x offset in **grid units** from the text-draw origin
    /// (`rect.x + text_offset_x - scroll_offset`).  Uses the same buffer config as
    /// the renderer so glyph positions are guaranteed to match.
    ///
    /// For left-aligned text pass `rect_w` = 0.0 (`buf_w` will be the window width).
    pub(crate) fn measure_cursor(
        &mut self,
        text: &str,
        font_size_grid: f32,
        rect_w_grid: f32,
        align: TextAlign,
        pw: f32,
        ph: f32,
        cursor_byte: usize,
        bold: bool,
        italic: bool,
        font: Option<&str>,
    ) -> f32 {
        let unit = ph / 1080.0;
        let font_size_px = font_size_grid * unit;
        // Match build_buffer: left-align uses full window width, others use rect width.
        let buf_w_px = match align {
            TextAlign::Left => pw,
            _ => (rect_w_grid * unit).max(1.0),
        };
        self.gpu_mut().text.measure_cursor_px(
            text,
            font_size_px,
            buf_w_px,
            ph,
            align,
            cursor_byte,
            bold,
            italic,
            font,
        ) / unit
    }

    /// Render into an existing encoder+view (overlay mode).
    pub(crate) fn render(
        &mut self,
        frame: GpuFrame<'_>,
        scene: &mut Scene,
        pw: f32,
        ph: f32,
        dt: f32,
        cursor: [f32; 3],
        clear_color: Option<crate::draw::Color>,
        tex_reg: &mut TextureRegistry,
    ) {
        let GpuFrame {
            encoder,
            view,
            device,
            queue,
        } = frame;
        self.advance_time(dt, pw, ph, cursor, queue, tex_reg);
        if scene.is_empty() {
            return;
        }
        let load_op = clear_color.map_or(wgpu::LoadOp::Load, |c| {
            wgpu::LoadOp::Clear(wgpu::Color {
                r: f64::from(c.r),
                g: f64::from(c.g),
                b: f64::from(c.b),
                a: f64::from(c.a),
            })
        });
        let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
            label: None,
            color_attachments: &[Some(wgpu::RenderPassColorAttachment {
                view,
                resolve_target: None,
                depth_slice: None,
                ops: wgpu::Operations {
                    load: load_op,
                    store: wgpu::StoreOp::Store,
                },
            })],
            depth_stencil_attachment: None,
            occlusion_query_set: None,
            timestamp_writes: None,
            multiview_mask: None,
        });
        // Build all geometry once; then interleave geometry+text per z-layer so
        // that text belonging to low-z items cannot bleed over high-z panels.
        let ranges = self.build_batches(scene, pw, ph, queue);
        scene.texts.sort_by(|a, b| a.z.total_cmp(&b.z));
        let split = scene.texts.partition_point(|t| t.z < 1.0);

        // Layer 0: normal widgets (z < 1.0)
        self.issue_draw_calls(&mut pass, pw, ph, &ranges, 0.0, 1.0, tex_reg);
        self.gpu_mut()
            .text
            .render(device, queue, &mut pass, &scene.texts[..split], pw, ph);

        // Layer 1: overlays — popouts, dropdowns (z >= 1.0)
        // Uses a separate TextSystem so the second prepare() doesn't overwrite
        // the vertex buffer that layer 0's render commands still reference.
        self.issue_draw_calls(&mut pass, pw, ph, &ranges, 1.0, f32::MAX, tex_reg);
        self.gpu_mut()
            .text_overlay
            .render(device, queue, &mut pass, &scene.texts[split..], pw, ph);
    }

    /// Present a frame (standalone mode). Acquires its own surface texture.
    pub(crate) fn present_standalone(
        &mut self,
        scene: &mut Scene,
        pw: f32,
        ph: f32,
        dt: f32,
        cursor: [f32; 3],
        clear_color: Option<crate::draw::Color>,
        tex_reg: &mut TextureRegistry,
    ) {
        let Some(sa) = &self.standalone else { return };
        // Advance time unconditionally so shader animations don't stall on empty frames.
        let queue = sa.queue.clone();
        self.advance_time(dt, pw, ph, cursor, &queue, tex_reg);
        if scene.is_empty() {
            return;
        }

        let sa = self.standalone.as_ref().unwrap();
        let output = match sa.surface.get_current_texture() {
            wgpu::CurrentSurfaceTexture::Success(t)
            | wgpu::CurrentSurfaceTexture::Suboptimal(t) => t,
            wgpu::CurrentSurfaceTexture::Lost | wgpu::CurrentSurfaceTexture::Outdated => {
                sa.surface.configure(&sa.device, &sa.config);
                return;
            }
            other => {
                eprintln!("[pane_ui] Surface error: {other:?}");
                return;
            }
        };

        let view = output
            .texture
            .create_view(&wgpu::TextureViewDescriptor::default());
        let device = sa.device.clone();
        let queue = sa.queue.clone();
        let mut enc = device.create_command_encoder(&wgpu::CommandEncoderDescriptor::default());
        self.render(
            GpuFrame {
                encoder: &mut enc,
                view: &view,
                device: &device,
                queue: &queue,
            },
            scene,
            pw,
            ph,
            dt,
            cursor,
            clear_color,
            tex_reg,
        );
        queue.submit(std::iter::once(enc.finish()));
        output.present();
    }

    /// Reconfigure the swapchain after a window resize (standalone mode only).
    fn resize(&mut self, w: u32, h: u32) {
        if let Some(sa) = &mut self.standalone {
            sa.config.width = w;
            sa.config.height = h;
            sa.surface.configure(&sa.device, &sa.config);
        }
    }

    /// Upload all batch vertices to the GPU and return per-batch draw metadata.
    /// Call `issue_draw_calls` one or more times to actually draw subsets.
    fn build_batches(
        &self,
        scene: &mut Scene,
        pw: f32,
        ph: f32,
        queue: &wgpu::Queue,
    ) -> Vec<BatchRange> {
        if scene.batches.is_empty() {
            return Vec::new();
        }
        scene.batches.sort_by(|a, b| a.z.total_cmp(&b.z));

        let mut all_verts: Vec<f32> = Vec::new();
        let mut ranges: Vec<BatchRange> = Vec::new();

        for batch in &scene.batches {
            let start = u32::try_from(all_verts.len() / 16).unwrap_or(0);

            let min_x = batch.verts.iter().map(|v| v[0]).fold(f32::MAX, f32::min);
            let max_x = batch.verts.iter().map(|v| v[0]).fold(f32::MIN, f32::max);
            let min_y = batch.verts.iter().map(|v| v[1]).fold(f32::MAX, f32::min);
            let max_y = batch.verts.iter().map(|v| v[1]).fold(f32::MIN, f32::max);
            let rw = (max_x - min_x).max(0.0001);
            let rh = (max_y - min_y).max(0.0001);
            let uv = batch.uv_rect.unwrap_or(UvRect::FULL);
            let unit = ph / 1080.0;
            let cr_px = (batch.corner_radius * unit).min(rw.min(rh) * unit * 0.5);
            let bw_px = batch.border_width * unit;
            let size_w = rw * unit;
            let size_h = rh * unit;
            let s = batch.state;

            for v in &batch.verts {
                let [x, y] = to_ndc(v[0], v[1], pw, ph);
                let t_u = (v[0] - min_x) / rw;
                let t_v = (v[1] - min_y) / rh;
                let u = uv.u_min + t_u * (uv.u_max - uv.u_min);
                let vv = uv.v_min + t_v * (uv.v_max - uv.v_min);
                all_verts.extend_from_slice(&[
                    x,
                    y,
                    u,
                    vv,
                    batch.color.r,
                    batch.color.g,
                    batch.color.b,
                    batch.color.a,
                    cr_px,
                    bw_px,
                    size_w,
                    size_h,
                    s[0],
                    s[1],
                    s[2],
                    s[3],
                ]);
            }
            ranges.push((
                batch.shader,
                start..u32::try_from(all_verts.len() / 16).unwrap_or(0),
                batch.clip,
                batch.texture,
                batch.z,
            ));
        }

        // Guard before writing — writing past the allocated buffer would be UB.
        if all_verts.len() > (MAX_VERTS * 16) as usize {
            eprintln!(
                "[pane] vertex buffer overflow: {} verts (max {}); frame skipped",
                all_verts.len() / 16,
                MAX_VERTS
            );
            return Vec::new();
        }
        queue.write_buffer(&self.gpu().vertex_buf, 0, bytemuck::cast_slice(&all_verts));
        ranges
    }

    /// Issue draw calls for all batches whose z is in `[z_min, z_max)`.
    fn issue_draw_calls(
        &self,
        pass: &mut wgpu::RenderPass,
        pw: f32,
        ph: f32,
        ranges: &[BatchRange],
        z_min: f32,
        z_max: f32,
        tex_reg: &TextureRegistry,
    ) {
        if ranges
            .iter()
            .all(|(_, _, _, _, z)| *z < z_min || *z >= z_max)
        {
            return;
        }
        // Re-bind our vertex buffer — glyphon may have replaced it during the text pass.
        pass.set_vertex_buffer(0, self.gpu().vertex_buf.slice(..));
        pass.set_bind_group(1, &self.gpu().globals_bg, &[]);
        let mut current_clip: Option<ClipRect> = None;
        for (shader_id, range, clip, texture, z) in ranges {
            if *z < z_min || *z >= z_max {
                continue;
            }
            if *clip != current_clip {
                match clip {
                    Some(c) => {
                        let (l, t, r, b) = clip_to_pixels(*c, pw, ph);
                        let max_w = pw.max(0.0) as u32;
                        let max_h = ph.max(0.0) as u32;
                        let sx = (l.max(0.0) as u32).min(max_w);
                        let sy = (t.max(0.0) as u32).min(max_h);
                        let sw = ((r - l).max(0.0) as u32).min(max_w - sx);
                        let sh = ((b - t).max(0.0) as u32).min(max_h - sy);
                        if sw > 0 && sh > 0 {
                            pass.set_scissor_rect(sx, sy, sw, sh);
                        }
                    }
                    None => pass.set_scissor_rect(0, 0, pw.max(0.0) as u32, ph.max(0.0) as u32),
                }
                current_clip = *clip;
            }
            pass.set_pipeline(&self.pipelines[shader_id.index()]);
            if let Some(id) = texture
                && tex_reg.is_hidden(*id)
            {
                continue;
            }
            let bg = texture.map_or_else(|| tex_reg.dummy(), |id| tex_reg.current_bind_group(id));
            pass.set_bind_group(0, bg, &[]);
            pass.draw(range.clone(), 0..1);
        }
        pass.set_scissor_rect(0, 0, pw.max(0.0) as u32, ph.max(0.0) as u32);
    }
}

// ── App (standalone event loop) ───────────────────────────────────────────────

struct App<F: FnMut(&mut Pane, &mut Scene, &mut Input, f32, f32, f32)> {
    window: Option<Arc<Window>>,
    pane: Pane,
    scene: Scene,
    input: Input,
    draw_fn: F,
    last_frame: std::time::Instant,
}

impl<F: FnMut(&mut Pane, &mut Scene, &mut Input, f32, f32, f32) + 'static> ApplicationHandler
    for App<F>
{
    fn resumed(&mut self, el: &ActiveEventLoop) {
        let win = Arc::new(
            el.create_window(Window::default_attributes().with_title("Pane"))
                .unwrap(),
        );
        if let Err(e) = pollster::block_on(self.pane.init(win.clone())) {
            eprintln!("{e}");
            el.exit();
            return;
        }
        self.window = Some(win);
    }

    fn window_event(&mut self, el: &ActiveEventLoop, _: WindowId, event: WindowEvent) {
        if let Some(win) = &self.window {
            let s = win.inner_size();
            self.input
                .handle_event(&event, s.width as f32, s.height as f32);
        }
        match event {
            WindowEvent::CloseRequested => el.exit(),
            WindowEvent::Resized(s) => self.pane.resize(s.width, s.height),
            WindowEvent::RedrawRequested => {
                if let Some(win) = &self.window {
                    let s = win.inner_size();
                    let pw = s.width as f32;
                    let ph = s.height as f32;
                    let now = std::time::Instant::now();
                    let dt = now.duration_since(self.last_frame).as_secs_f32().min(0.1);
                    self.last_frame = now;
                    self.scene.clear();
                    (self.draw_fn)(&mut self.pane, &mut self.scene, &mut self.input, pw, ph, dt);
                }
            }
            _ => {}
        }
    }

    fn about_to_wait(&mut self, _: &ActiveEventLoop) {
        self.input.poll_gamepad();
        if let Some(win) = &self.window {
            win.request_redraw();
        }
    }
}

// ── Entry Point ───────────────────────────────────────────────────────────────

/// Run a standalone pane application, blocking until the window is closed.
///
/// `draw_fn` is called every frame with `(pane, scene, input, pw, ph, dt)` where
/// `pw`/`ph` are the window pixel dimensions and `dt` is the frame delta in seconds.
/// The closure is responsible for calling `run_frame` which drives
/// widget logic, fills the scene, and presents the frame.
///
/// # Panics
///
/// Panics if the event loop or window cannot be created.
pub fn run<F: FnMut(&mut Pane, &mut Scene, &mut Input, f32, f32, f32) + 'static>(draw_fn: F) {
    let el = EventLoop::new().unwrap();
    let mut app = App {
        window: None,
        pane: Pane::new_standalone(),
        scene: Scene::new(),
        input: Input::new_with_gilrs(),
        draw_fn,
        last_frame: std::time::Instant::now(),
    };
    el.run_app(&mut app).unwrap();
}