/**
 * WebGL2 sprite renderer: one indexed quad batcher with per-quad tint that can
 * sample several atlas pages inside a single draw call.
 *
 * Diablo-style rendering is a flood of axis-aligned quads — floors, walls,
 * units, items — drawn in painter's order. So the renderer accumulates
 * world-space quads (position, source rectangle, tint) and submits them in as
 * few calls as it can. Depth ordering is the caller's job (draw order), not the
 * GPU's, which is how the original games did it and what keeps the batcher
 * trivial.
 *
 * Three properties keep the batch big and the CPU cost low:
 *
 * - **Multi-texture batching.** A Diablo level needs more tiles than one 2048²
 *   page holds, and painter's order interleaves those pages constantly: sorting
 *   by texture is not allowed, because that is exactly the order that makes the
 *   scene correct. Binding one texture per batch therefore used to split the
 *   frame into hundreds of draw calls (measured: up to 3077 on
 *   `act4/107-act-4-lava-1`). Instead every page in flight gets its own texture
 *   unit, the unit index rides along as a vertex attribute, and the fragment
 *   shader picks the sampler — so painter's order survives *and* the whole frame
 *   is usually one `drawElements`. Only when a batch needs more distinct pages
 *   than the hardware has units does it flush and start over.
 * - **Indexed geometry.** Four vertices per quad instead of six, with a static
 *   element buffer, so a third less vertex data crosses the bus per frame.
 * - **Zero-allocation vertex writes.** {@link SpriteRenderer.quad} writes the
 *   vertex floats straight out, unrolled. The earlier version built four
 *   throwaway arrays per quad, which at ~2800 quads a frame meant ~11k
 *   short-lived objects per frame handed to the GC.
 *
 * Tint is a vertex attribute rather than a uniform, so coloured geometry
 * (a floor rectangle, a wall) batches together with sprites instead of forcing
 * a flush per colour.
 */
import type { AtlasFrame, SpriteAtlas } from './atlas.ts'

/** Camera state for one frame. */
export interface Camera {
  /** World-space centre of the view. */
  readonly x: number
  /** World-space centre of the view. */
  readonly y: number
  /** Pixels per world unit. */
  readonly zoom: number
}

/**
 * One uploaded atlas texture.
 *
 * Diablo II maps need more tiles than a single 2048² page holds, so the renderer
 * keeps several textures. Each quad records which page it samples; the batcher
 * assigns pages to texture units so a page switch no longer splits the batch
 * (see {@link SpriteRenderer.draw}).
 */
export interface AtlasHandle {
  /** The GPU texture. */
  readonly texture: WebGLTexture
  /** Texture width in pixels, for UV maths. */
  readonly width: number
  /** Texture height in pixels. */
  readonly height: number
}

/** Draw options for one sprite quad. */
export interface DrawOptions {
  /** Multiplicative tint, defaulting to opaque white. */
  readonly tint?: readonly [number, number, number, number]
  /** Mirror horizontally. */
  readonly flipX?: boolean
  /** Atlas page to sample; defaults to the most recently set atlas. */
  readonly atlas?: AtlasHandle
}

/** Configuration options and event callbacks for SpriteRenderer. */
export interface RendererOptions {
  /** Optional callback invoked when the WebGL context is lost. */
  readonly onContextLost?: (event: Event) => void
  /** Optional callback invoked when the WebGL context is restored. */
  readonly onContextRestored?: () => void
}

/** Vertices per quad: the four corners, reused by the index buffer. */
const VERTICES_PER_QUAD = 4
/** Indices per quad (two triangles). */
const INDICES_PER_QUAD = 6
/** Floats per vertex: x, y, u, v, r, g, b, a, unit. */
const FLOATS_PER_VERTEX = 9
/** Initial batch capacity in quads. */
const INITIAL_CAPACITY = 8192
/**
 * Texture units the batcher will use at most.
 *
 * WebGL2 guarantees `MAX_TEXTURE_IMAGE_UNITS >= 16`, and the packs top out at
 * ten pages for a level (`act5/109-act-5-town-townwest`) plus one character
 * atlas, so sixteen is enough to hold an entire frame's pages at once. The real
 * limit is still queried at construction and respected.
 */
const MAX_BATCH_TEXTURES = 16

/**
 * Build the vertex shader.
 *
 * @returns GLSL ES 3.00 source.
 */
function vertexShaderSource(): string {
  return `#version 300 es
in vec2 a_position;
in vec2 a_uv;
in vec4 a_tint;
in float a_unit;
uniform vec2 u_camera;
uniform vec2 u_viewport;
uniform float u_zoom;
out vec2 v_uv;
out vec4 v_tint;
flat out int v_unit;
void main() {
  vec2 offset = (a_position - u_camera) * u_zoom;
  // 'half' is a reserved word in GLSL ES 3.00, hence the longer name.
  vec2 halfViewport = u_viewport * 0.5;
  gl_Position = vec4(offset.x / halfViewport.x, -offset.y / halfViewport.y, 0.0, 1.0);
  v_uv = a_uv;
  v_tint = a_tint;
  v_unit = int(a_unit);
}`
}

/**
 * Build the fragment shader for a given texture-unit count.
 *
 * GLSL ES 3.00 forbids indexing a sampler array with a non-constant expression,
 * so the sampler is picked by a `switch` over constant cases. The case list is
 * generated to match however many units the context actually offers.
 *
 * @param units - number of samplers to expose.
 * @returns GLSL ES 3.00 source.
 */
function fragmentShaderSource(units: number): string {
  const cases: string[] = []
  for (let unit = 1; unit < units; unit += 1) {
    cases.push(`    case ${String(unit)}: texel = texture(u_textures[${String(unit)}], v_uv); break;`)
  }
  return `#version 300 es
precision mediump float;
in vec2 v_uv;
in vec4 v_tint;
flat in int v_unit;
uniform sampler2D u_textures[${String(units)}];
out vec4 outColor;
void main() {
  vec4 texel;
  switch (v_unit) {
${cases.join('\n')}
    default: texel = texture(u_textures[0], v_uv); break;
  }
  if (texel.a == 0.0) discard;
  outColor = texel * v_tint;
}`
}

/** Raised when the context or program cannot be created. */
export class RendererError extends Error {}

/**
 * A quad batcher bound to one canvas.
 */
export class SpriteRenderer {
  /** The target canvas. */
  readonly canvas: HTMLCanvasElement
  /** The WebGL2 context, exposed for the few callers that need raw access. */
  readonly gl: WebGL2RenderingContext
  private readonly program: WebGLProgram
  private readonly vao: WebGLVertexArrayObject
  private readonly buffer: WebGLBuffer
  private readonly indexBuffer: WebGLBuffer
  private readonly atlasTexture: WebGLTexture
  private readonly whiteTexture: WebGLTexture
  /** The 1×1 white page used by `drawSolid`. */
  private readonly whiteAtlas: AtlasHandle
  /** The default page (what `setAtlas` uploads into). */
  private defaultAtlas: AtlasHandle
  private readonly uniformCamera: WebGLUniformLocation
  private readonly uniformViewport: WebGLUniformLocation
  private readonly uniformZoom: WebGLUniformLocation
  private readonly uniformTextures: WebGLUniformLocation
  private vertices: Float32Array
  private indices: Uint32Array
  private quadCount = 0
  private atlasWidth = 1
  private atlasHeight = 1
  /** Texture units this context offers the batcher. */
  private readonly maxBatchTextures: number
  /** Pages bound to units `0..batchTextureCount-1` for the pending batch. */
  private readonly batchTextures: (AtlasHandle | null)[]
  private batchTextureCount = 0
  private camera: Camera = { x: 0, y: 0, zoom: 1 }

  /** `drawElements` calls issued since the last {@link begin}. */
  private frameDrawCalls = 0
  /** Quads submitted since the last {@link begin}. */
  private frameQuads = 0

  /** All textures allocated and managed by this renderer. */
  private readonly allocatedTextures = new Set<WebGLTexture>()
  /** Whether this renderer has been disposed. */
  private disposed = false
  /** Listener references for cleanup. */
  private readonly handleContextLost: (event: Event) => void
  private readonly handleContextRestored: () => void

  /**
   * @param canvas - the canvas to render into.
   * @param options - optional lifecycle hooks and configuration.
   */
  constructor(canvas: HTMLCanvasElement, options: RendererOptions = {}) {
    this.canvas = canvas
    const gl = canvas.getContext('webgl2', { alpha: false, antialias: false, premultipliedAlpha: false })
    if (gl === null) throw new RendererError('WebGL2 is not available in this browser')
    this.gl = gl

    const reportedUnits = gl.getParameter(gl.MAX_TEXTURE_IMAGE_UNITS) as number | null
    // WebGL2 guarantees at least 16; clamp to at least 1 so a stubbed context
    // (the lifecycle tests) still produces a valid shader.
    const availableUnits = typeof reportedUnits === 'number' && reportedUnits > 0 ? reportedUnits : MAX_BATCH_TEXTURES
    this.maxBatchTextures = Math.max(1, Math.min(MAX_BATCH_TEXTURES, availableUnits))
    this.batchTextures = new Array<AtlasHandle | null>(this.maxBatchTextures).fill(null)

    this.program = createProgram(gl, vertexShaderSource(), fragmentShaderSource(this.maxBatchTextures))
    const vao = gl.createVertexArray()
    const buffer = gl.createBuffer()
    const indexBuffer = gl.createBuffer()
    const atlasTexture = gl.createTexture()
    const whiteTexture = gl.createTexture()
    if (vao === null || buffer === null || indexBuffer === null || atlasTexture === null || whiteTexture === null) {
      throw new RendererError('WebGL2 resource allocation failed')
    }
    this.vao = vao
    this.buffer = buffer
    this.indexBuffer = indexBuffer
    this.atlasTexture = atlasTexture
    this.whiteTexture = whiteTexture
    this.allocatedTextures.add(atlasTexture)
    this.allocatedTextures.add(whiteTexture)
    this.defaultAtlas = { texture: atlasTexture, width: 1, height: 1 }
    this.whiteAtlas = { texture: whiteTexture, width: 1, height: 1 }
    this.vertices = new Float32Array(INITIAL_CAPACITY * VERTICES_PER_QUAD * FLOATS_PER_VERTEX)
    this.indices = buildQuadIndices(INITIAL_CAPACITY)

    const attributePosition = gl.getAttribLocation(this.program, 'a_position')
    const attributeUv = gl.getAttribLocation(this.program, 'a_uv')
    const attributeTint = gl.getAttribLocation(this.program, 'a_tint')
    const attributeUnit = gl.getAttribLocation(this.program, 'a_unit')
    gl.bindVertexArray(vao)
    gl.bindBuffer(gl.ARRAY_BUFFER, buffer)
    gl.bufferData(gl.ARRAY_BUFFER, this.vertices.byteLength, gl.DYNAMIC_DRAW)
    const stride = FLOATS_PER_VERTEX * 4
    gl.enableVertexAttribArray(attributePosition)
    gl.vertexAttribPointer(attributePosition, 2, gl.FLOAT, false, stride, 0)
    gl.enableVertexAttribArray(attributeUv)
    gl.vertexAttribPointer(attributeUv, 2, gl.FLOAT, false, stride, 8)
    gl.enableVertexAttribArray(attributeTint)
    gl.vertexAttribPointer(attributeTint, 4, gl.FLOAT, false, stride, 16)
    // A stubbed context can report -1 for an attribute it does not know about;
    // enabling that location would be an INVALID_VALUE.
    if (attributeUnit >= 0) {
      gl.enableVertexAttribArray(attributeUnit)
      gl.vertexAttribPointer(attributeUnit, 1, gl.FLOAT, false, stride, 32)
    }
    gl.bindBuffer(gl.ELEMENT_ARRAY_BUFFER, indexBuffer)
    gl.bufferData(gl.ELEMENT_ARRAY_BUFFER, this.indices, gl.STATIC_DRAW)
    gl.bindVertexArray(null)

    this.uniformCamera = requireUniform(gl, this.program, 'u_camera')
    this.uniformViewport = requireUniform(gl, this.program, 'u_viewport')
    this.uniformZoom = requireUniform(gl, this.program, 'u_zoom')
    this.uniformTextures = requireUniform(gl, this.program, 'u_textures[0]')

    // The sampler array never changes: unit i always reads texture unit i.
    const samplerUnits = new Int32Array(this.maxBatchTextures)
    for (let unit = 0; unit < this.maxBatchTextures; unit += 1) samplerUnits[unit] = unit
    gl.useProgram(this.program)
    gl.uniform1iv(this.uniformTextures, samplerUnits)

    for (const texture of [atlasTexture, whiteTexture]) {
      gl.bindTexture(gl.TEXTURE_2D, texture)
      gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_MIN_FILTER, gl.NEAREST)
      gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_MAG_FILTER, gl.NEAREST)
      gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_WRAP_S, gl.CLAMP_TO_EDGE)
      gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_WRAP_T, gl.CLAMP_TO_EDGE)
    }
    gl.bindTexture(gl.TEXTURE_2D, whiteTexture)
    gl.texImage2D(gl.TEXTURE_2D, 0, gl.RGBA, 1, 1, 0, gl.RGBA, gl.UNSIGNED_BYTE, new Uint8Array([255, 255, 255, 255]))
    gl.enable(gl.BLEND)
    gl.blendFunc(gl.SRC_ALPHA, gl.ONE_MINUS_SRC_ALPHA)
    gl.disable(gl.DEPTH_TEST)

    this.handleContextLost = (event: Event): void => {
      event.preventDefault()
      console.warn('WebGL2 context lost')
      options.onContextLost?.(event)
    }
    this.handleContextRestored = (): void => {
      console.info('WebGL2 context restored')
      options.onContextRestored?.()
    }
    canvas.addEventListener('webglcontextlost', this.handleContextLost)
    canvas.addEventListener('webglcontextrestored', this.handleContextRestored)
  }

  /** Whether this renderer and its GPU resources have been disposed. */
  get isDisposed(): boolean {
    return this.disposed
  }

  /**
   * Draw calls issued since the last {@link begin}.
   *
   * This is the real measured count, not an estimate: the HUD used to report a
   * hardcoded `1` while the frame was actually spending hundreds of calls, which
   * is precisely why the cost stayed invisible.
   */
  get drawCalls(): number {
    return this.frameDrawCalls
  }

  /** Quads submitted since the last {@link begin}. */
  get quadsSubmitted(): number {
    return this.frameQuads
  }

  /** Texture units this batcher can bind at once. */
  get textureUnits(): number {
    return this.maxBatchTextures
  }

  /**
   * Upload an atlas, replacing the current one.
   *
   * @param atlas - the packed atlas.
   * @returns the handle other pages can be drawn with.
   */
  setAtlas(atlas: SpriteAtlas): AtlasHandle {
    if (this.disposed) throw new RendererError('SpriteRenderer has already been disposed')
    const gl = this.gl
    // Uploading past the limit is a silent INVALID_VALUE in GL; say what
    // actually went wrong instead.
    const limit = gl.getParameter(gl.MAX_TEXTURE_SIZE) as number
    if (atlas.width > limit || atlas.height > limit) {
      throw new RendererError(
        `atlas ${String(atlas.width)}x${String(atlas.height)} exceeds the ${String(limit)}px texture limit`,
      )
    }
    this.flush()
    gl.bindTexture(gl.TEXTURE_2D, this.atlasTexture)
    gl.pixelStorei(gl.UNPACK_ALIGNMENT, 1)
    gl.texImage2D(
      gl.TEXTURE_2D, 0, gl.RGBA, atlas.width, atlas.height, 0,
      gl.RGBA, gl.UNSIGNED_BYTE, new Uint8Array(atlas.pixels.buffer, atlas.pixels.byteOffset, atlas.pixels.byteLength),
    )
    this.atlasWidth = Math.max(atlas.width, 1)
    this.atlasHeight = Math.max(atlas.height, 1)
    this.defaultAtlas = { texture: this.atlasTexture, width: this.atlasWidth, height: this.atlasHeight }
    return this.defaultAtlas
  }

  /**
   * Upload an extra atlas page and return a handle for it.
   *
   * Packed maps ship several pages; the page can load the ones the spawn area
   * needs first and hand them over as they arrive, instead of waiting for one
   * gigantic texture to decode.
   *
   * @param source - decoded pixels or an `ImageBitmap` (PNG straight from cache).
   * @param width - page width in pixels.
   * @param height - page height in pixels.
   * @returns the handle to pass to {@link draw}.
   */
  addAtlas(
    source: ImageBitmap | { pixels: Uint8ClampedArray | Uint8Array; width: number; height: number },
    width?: number,
    height?: number,
  ): AtlasHandle {
    if (this.disposed) throw new RendererError('SpriteRenderer has already been disposed')
    const gl = this.gl
    const limit = gl.getParameter(gl.MAX_TEXTURE_SIZE) as number
    const isBitmap = typeof ImageBitmap !== 'undefined' && source instanceof ImageBitmap
    const sourceWidth = isBitmap ? (source as ImageBitmap).width : (source as { width: number }).width
    const sourceHeight = isBitmap ? (source as ImageBitmap).height : (source as { height: number }).height
    const pageWidth = Math.max(1, width ?? sourceWidth)
    const pageHeight = Math.max(1, height ?? sourceHeight)
    if (pageWidth > limit || pageHeight > limit) {
      throw new RendererError(`atlas ${String(pageWidth)}x${String(pageHeight)} exceeds the ${String(limit)}px texture limit`)
    }
    this.flush()
    const texture = gl.createTexture()
    if (texture === null) throw new RendererError('texture allocation failed')
    this.allocatedTextures.add(texture)
    gl.bindTexture(gl.TEXTURE_2D, texture)
    gl.pixelStorei(gl.UNPACK_ALIGNMENT, 1)
    gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_MIN_FILTER, gl.NEAREST)
    gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_MAG_FILTER, gl.NEAREST)
    gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_WRAP_S, gl.CLAMP_TO_EDGE)
    gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_WRAP_T, gl.CLAMP_TO_EDGE)
    if (isBitmap) {
      gl.texImage2D(gl.TEXTURE_2D, 0, gl.RGBA, gl.RGBA, gl.UNSIGNED_BYTE, source as ImageBitmap)
    } else {
      const indexed = source as { pixels: Uint8ClampedArray | Uint8Array; width: number; height: number }
      const pixels = indexed.pixels instanceof Uint8Array
        ? indexed.pixels
        : new Uint8Array(indexed.pixels.buffer, indexed.pixels.byteOffset, indexed.pixels.byteLength)
      gl.texImage2D(
        gl.TEXTURE_2D, 0, gl.RGBA, indexed.width, indexed.height, 0,
        gl.RGBA, gl.UNSIGNED_BYTE, pixels,
      )
    }
    return { texture, width: pageWidth, height: pageHeight }
  }

  /**
   * Release an atlas texture handle and free its GPU memory.
   *
   * @param handle - the atlas handle previously returned by addAtlas or setAtlas.
   * @returns true if the texture was managed by this renderer and deleted, false otherwise.
   */
  deleteAtlas(handle: AtlasHandle): boolean {
    if (this.disposed || handle === this.whiteAtlas) return false
    if (!this.allocatedTextures.has(handle.texture)) return false

    // Quads already queued may reference this page's unit, so they have to go
    // out before the texture disappears.
    for (let unit = 0; unit < this.batchTextureCount; unit += 1) {
      if (this.batchTextures[unit] === handle) {
        this.flush()
        break
      }
    }
    if (this.defaultAtlas === handle) {
      this.defaultAtlas = this.whiteAtlas
    }

    if (!this.gl.isContextLost()) {
      this.gl.deleteTexture(handle.texture)
    }
    this.allocatedTextures.delete(handle.texture)
    return true
  }

  /** The page `setAtlas` uploaded into, for callers that draw with handles. */
  get defaultAtlasHandle(): AtlasHandle {
    return this.defaultAtlas
  }

  /**
   * Start a frame: clear, and record the camera.
   *
   * @param camera - camera to render with.
   * @param clear - background colour as `[r, g, b]` in 0..1.
   */
  begin(camera: Camera, clear: readonly [number, number, number] = [0, 0, 0]): void {
    if (this.disposed) throw new RendererError('SpriteRenderer has already been disposed')
    const gl = this.gl
    this.camera = camera
    this.quadCount = 0
    this.batchTextureCount = 0
    this.frameDrawCalls = 0
    this.frameQuads = 0
    gl.viewport(0, 0, gl.drawingBufferWidth, gl.drawingBufferHeight)
    gl.clearColor(clear[0], clear[1], clear[2], 1)
    gl.clear(gl.COLOR_BUFFER_BIT)
  }

  /**
   * Queue one atlas sprite.
   *
   * @param frame - the atlas placement to draw.
   * @param x - world-space left edge.
   * @param y - world-space top edge.
   * @param options - tint and flipping.
   */
  draw(frame: AtlasFrame, x: number, y: number, options: DrawOptions = {}): void {
    if (this.disposed) return
    const page = options.atlas ?? this.defaultAtlas
    const unit = this.unitFor(page)
    const u0 = frame.x / page.width
    const v0 = frame.y / page.height
    const u1 = (frame.x + frame.width) / page.width
    const v1 = (frame.y + frame.height) / page.height
    const tint = options.tint
    const flipped = options.flipX === true
    this.quad(
      x, y, x + frame.width, y + frame.height,
      flipped ? u1 : u0, v0,
      flipped ? u0 : u1, v1,
      tint === undefined ? 1 : tint[0],
      tint === undefined ? 1 : tint[1],
      tint === undefined ? 1 : tint[2],
      tint === undefined ? 1 : tint[3],
      unit,
    )
  }

  /**
   * Queue one solid-colour rectangle.
   *
   * @param x - world-space left edge.
   * @param y - world-space top edge.
   * @param width - rectangle width.
   * @param height - rectangle height.
   * @param color - `[r, g, b, a]` in 0..1.
   */
  drawSolid(x: number, y: number, width: number, height: number, color: readonly [number, number, number, number]): void {
    if (this.disposed) return
    const unit = this.unitFor(this.whiteAtlas)
    this.quad(x, y, x + width, y + height, 0.5, 0.5, 0.5, 0.5, color[0], color[1], color[2], color[3], unit)
  }

  /**
   * Find the texture unit a page is bound to for the pending batch, binding it
   * to a free unit if this is its first quad.
   *
   * Running out of units is the only thing that still splits a batch, and with
   * sixteen of them against a ten-page worst case that effectively never
   * happens for a map.
   *
   * @param page - the atlas page the caller wants to sample.
   * @returns the texture unit index to write into the vertices.
   */
  private unitFor(page: AtlasHandle): number {
    const count = this.batchTextureCount
    for (let unit = 0; unit < count; unit += 1) {
      if (this.batchTextures[unit] === page) return unit
    }
    if (count === this.maxBatchTextures) {
      this.flush()
      this.batchTextures[0] = page
      this.batchTextureCount = 1
      return 0
    }
    this.batchTextures[count] = page
    this.batchTextureCount = count + 1
    return count
  }

  /**
   * Append one quad's four vertices.
   *
   * Written out longhand on purpose: the readable loop-over-temporary-arrays
   * version allocated four arrays per quad, and at a few thousand quads a frame
   * that dominated the CPU side of the frame (measured 2.3× slower).
   *
   * @param x0 - left edge.
   * @param y0 - top edge.
   * @param x1 - right edge.
   * @param y1 - bottom edge.
   * @param u0 - left texture coordinate.
   * @param v0 - top texture coordinate.
   * @param u1 - right texture coordinate.
   * @param v1 - bottom texture coordinate.
   * @param r - tint red.
   * @param g - tint green.
   * @param b - tint blue.
   * @param a - tint alpha.
   * @param unit - texture unit to sample.
   */
  private quad(
    x0: number, y0: number, x1: number, y1: number,
    u0: number, v0: number, u1: number, v1: number,
    r: number, g: number, b: number, a: number,
    unit: number,
  ): void {
    const needed = (this.quadCount + 1) * VERTICES_PER_QUAD * FLOATS_PER_VERTEX
    if (needed > this.vertices.length) this.grow(needed)
    const v = this.vertices
    let at = this.quadCount * VERTICES_PER_QUAD * FLOATS_PER_VERTEX
    // Corner 0: top-left.
    v[at] = x0; v[at + 1] = y0; v[at + 2] = u0; v[at + 3] = v0
    v[at + 4] = r; v[at + 5] = g; v[at + 6] = b; v[at + 7] = a; v[at + 8] = unit
    at += FLOATS_PER_VERTEX
    // Corner 1: top-right.
    v[at] = x1; v[at + 1] = y0; v[at + 2] = u1; v[at + 3] = v0
    v[at + 4] = r; v[at + 5] = g; v[at + 6] = b; v[at + 7] = a; v[at + 8] = unit
    at += FLOATS_PER_VERTEX
    // Corner 2: bottom-left.
    v[at] = x0; v[at + 1] = y1; v[at + 2] = u0; v[at + 3] = v1
    v[at + 4] = r; v[at + 5] = g; v[at + 6] = b; v[at + 7] = a; v[at + 8] = unit
    at += FLOATS_PER_VERTEX
    // Corner 3: bottom-right.
    v[at] = x1; v[at + 1] = y1; v[at + 2] = u1; v[at + 3] = v1
    v[at + 4] = r; v[at + 5] = g; v[at + 6] = b; v[at + 7] = a; v[at + 8] = unit
    this.quadCount += 1
    this.frameQuads += 1
  }

  /** Submit every queued quad. */
  flush(): void {
    const gl = this.gl
    if (this.disposed || this.quadCount === 0) return
    const floatCount = this.quadCount * VERTICES_PER_QUAD * FLOATS_PER_VERTEX
    gl.useProgram(this.program)
    gl.bindVertexArray(this.vao)
    gl.bindBuffer(gl.ARRAY_BUFFER, this.buffer)
    // Orphan first: without it the driver has to wait until the GPU is done
    // reading the range this overwrites, which is a pipeline stall on every
    // flush. Handing back a fresh block lets both sides run.
    gl.bufferData(gl.ARRAY_BUFFER, this.vertices.byteLength, gl.DYNAMIC_DRAW)
    gl.bufferSubData(gl.ARRAY_BUFFER, 0, this.vertices, 0, floatCount)
    gl.uniform2f(this.uniformCamera, this.camera.x, this.camera.y)
    gl.uniform2f(this.uniformViewport, gl.drawingBufferWidth, gl.drawingBufferHeight)
    gl.uniform1f(this.uniformZoom, this.camera.zoom)
    for (let unit = 0; unit < this.batchTextureCount; unit += 1) {
      const page = this.batchTextures[unit]
      if (page === undefined || page === null) continue
      gl.activeTexture(gl.TEXTURE0 + unit)
      gl.bindTexture(gl.TEXTURE_2D, page.texture)
    }
    gl.drawElements(gl.TRIANGLES, this.quadCount * INDICES_PER_QUAD, gl.UNSIGNED_INT, 0)
    this.frameDrawCalls += 1
    this.quadCount = 0
    this.batchTextureCount = 0
  }

  /**
   * Destroy all WebGL2 GPU resources (textures, VAO, VBO, program) and detach event listeners.
   */
  dispose(): void {
    if (this.disposed) return
    this.disposed = true

    this.canvas.removeEventListener('webglcontextlost', this.handleContextLost)
    this.canvas.removeEventListener('webglcontextrestored', this.handleContextRestored)

    const gl = this.gl
    if (!gl.isContextLost()) {
      for (const texture of this.allocatedTextures) {
        gl.deleteTexture(texture)
      }
      this.allocatedTextures.clear()
      gl.deleteBuffer(this.buffer)
      gl.deleteBuffer(this.indexBuffer)
      gl.deleteVertexArray(this.vao)
      gl.deleteProgram(this.program)
    } else {
      this.allocatedTextures.clear()
    }
    this.quadCount = 0
    this.batchTextureCount = 0
  }

  /**
   * Grow the vertex and index buffers, preserving queued geometry.
   *
   * @param needed - required float count.
   */
  private grow(needed: number): void {
    let capacity = this.vertices.length
    while (capacity < needed) capacity *= 2
    const grown = new Float32Array(capacity)
    grown.set(this.vertices)
    this.vertices = grown
    const quads = Math.floor(capacity / (VERTICES_PER_QUAD * FLOATS_PER_VERTEX))
    this.indices = buildQuadIndices(quads)
    const gl = this.gl
    gl.bindVertexArray(this.vao)
    gl.bindBuffer(gl.ARRAY_BUFFER, this.buffer)
    gl.bufferData(gl.ARRAY_BUFFER, grown.byteLength, gl.DYNAMIC_DRAW)
    gl.bindBuffer(gl.ELEMENT_ARRAY_BUFFER, this.indexBuffer)
    gl.bufferData(gl.ELEMENT_ARRAY_BUFFER, this.indices, gl.STATIC_DRAW)
    gl.bindVertexArray(null)
  }
}

/**
 * Build the static index buffer contents for a quad capacity.
 *
 * Every quad is two triangles over its four corners: top-left, top-right,
 * bottom-left and bottom-right, wound the same way the old six-vertex layout
 * was so nothing about the rasterised result changes.
 *
 * @param quads - number of quads to cover.
 * @returns indices, `6 * quads` long.
 */
function buildQuadIndices(quads: number): Uint32Array {
  const indices = new Uint32Array(quads * INDICES_PER_QUAD)
  for (let quad = 0; quad < quads; quad += 1) {
    const vertex = quad * VERTICES_PER_QUAD
    const at = quad * INDICES_PER_QUAD
    indices[at] = vertex
    indices[at + 1] = vertex + 1
    indices[at + 2] = vertex + 2
    indices[at + 3] = vertex + 1
    indices[at + 4] = vertex + 3
    indices[at + 5] = vertex + 2
  }
  return indices
}

/**
 * Compile and link a program, reporting shader logs on failure.
 *
 * Shaders are detached and flagged for deletion immediately after linking to
 * prevent driver-level memory leaks.
 *
 * @param gl - the context.
 * @param vertexSource - vertex shader source.
 * @param fragmentSource - fragment shader source.
 * @returns the linked program.
 */
function createProgram(gl: WebGL2RenderingContext, vertexSource: string, fragmentSource: string): WebGLProgram {
  const compile = (type: number, source: string): WebGLShader => {
    const shader = gl.createShader(type)
    if (shader === null) throw new RendererError('could not create shader')
    gl.shaderSource(shader, source)
    gl.compileShader(shader)
    if (gl.getShaderParameter(shader, gl.COMPILE_STATUS) !== true) {
      const log = gl.getShaderInfoLog(shader) ?? 'unknown'
      gl.deleteShader(shader)
      throw new RendererError(`shader compile failed: ${log}`)
    }
    return shader
  }
  const program = gl.createProgram()
  if (program === null) throw new RendererError('could not create program')
  const vs = compile(gl.VERTEX_SHADER, vertexSource)
  const fs = compile(gl.FRAGMENT_SHADER, fragmentSource)
  gl.attachShader(program, vs)
  gl.attachShader(program, fs)
  gl.linkProgram(program)
  const linked = gl.getProgramParameter(program, gl.LINK_STATUS) === true
  const log = linked ? '' : (gl.getProgramInfoLog(program) ?? 'unknown')
  gl.detachShader(program, vs)
  gl.deleteShader(vs)
  gl.detachShader(program, fs)
  gl.deleteShader(fs)
  if (!linked) {
    gl.deleteProgram(program)
    throw new RendererError(`program link failed: ${log}`)
  }
  return program
}

/**
 * Fetch a uniform location or fail loudly.
 *
 * @param gl - the context.
 * @param program - the linked program.
 * @param name - uniform name.
 * @returns the location.
 */
function requireUniform(gl: WebGL2RenderingContext, program: WebGLProgram, name: string): WebGLUniformLocation {
  const location = gl.getUniformLocation(program, name)
  if (location === null) throw new RendererError(`uniform ${name} is missing from the program`)
  return location
}