v0.0.1

Added

• upload-once / draw-many GPU API for interactive viewers
• wgpu headless GPU renderer — scanline-parity shading modes, framing, SSAA

Added

• GpuPathTracer: a wgpu compute-shader port of oxideav-render's
  unbiased path tracer. It follows the CPU estimator spec
  (oxideav_render::pathtrace §1–5): the same PCG / Owen-scrambled
  Sobol' sequences and dimension layout, NEE to punctual lights plus
  one emissive-triangle (Arvo spherical-triangle sampling) or
  environment sample, power-heuristic MIS, Russian roulette,
  stochastic BLEND coins, and the layered glTF BSDF (transmission,
  volume, clearcoat, sheen). Its images match the CPU tracer to float
  rounding (63–85 dB PSNR at 64 spp).

• Software BVH traversal in WGSL over TraceScene's SAH BVH, using
  four storage buffers. Textures go into an Rgba16Float array-texture
  atlas, filtered manually so the result matches the CPU sampler.

• Progressive API: sync / refine / samples / is_converged /
  image / hdr / draw_texture (no readback) / finish /
  set_environment / invalidate_scene / reset, with the CPU
  PathTracer's reset rules. It implements Renderer (render and
  render_hdr take samples_per_pixel samples).

• GpuMode and GpuRenderer::set_mode / mode / path_tracer. The
  path tracer is registered as "gpu-pathtrace"
  (PATHTRACE_BACKEND_NAME) next to "gpu".

• tests/pathtrace.rs (GPU-vs-CPU parity suite), the pt_bench
  example and BENCHMARKS.md. The Cornell box at 256², 64 spp renders
  in 32 ms on an RTX 5080, against 281 ms for the CPU tracer.

• GpuRenderer: wgpu-backed headless implementation of
  oxideav_render::Renderer (runtime-loaded Vulkan / Metal / DX12 / GL).
  Covers the scanline backend's shading modes (Flat, Gouraud, Phong,
  Wireframe, NormalDebug, DepthDebug), camera framing, and supersampled
  AA, plus hardware clipping and depth testing.

• GpuBackend API selector, GpuRenderer::from_device for sharing an
  application's wgpu device, adapter_summary().

• register_into / BACKEND_NAME ("gpu") for RenderRegistry.

• GPU-vs-scanline parity tests (skip without an adapter) and the
  gpu_vs_cpu example.

• GpuRenderer::upload / draw: keep a scene resident on the GPU and
  redraw it into an offscreen texture (COLOR_FORMAT, sRGB-encoded,
  COPY_SRC | TEXTURE_BINDING) without readback, for interactive
  viewers; device() / queue() accessors.

• Rebuilt on oxideav-render's scene-preparation layer: PreparedScene
  upload (posing at time, materials, textures, lights, cameras) and
  Camera::resolve framing.

• ShadingMode::Pbr: glTF 2.0 Appendix B metallic-roughness BRDF with
  all five core texture slots (uploaded as Rgba16Float mip chains),
  KHR_texture_transform, TEXCOORD_1, vertex colours, unlit, MASK /
  BLEND (sorted, linear compositing), double-sided culling, up to 16
  punctual lights, ambient, exposure and tone mapping.

• Two-pass frame: scene-linear Rgba16Float + coverage target, then a
  GPU resolve pass (background composite, supersample average, sRGB
  encode); draw now honours aa.

• set_texture_resolver.

• Shadow maps (opts.shadows) for up to 4 directional / spot lights:
  R32Float linear-depth array, same light-space fit and normal-offset
  3×3 bilinear PCF as the scanline backend, MASK casters honour their
  cutoff, BLEND surfaces cast nothing.

• GPU-vs-scanline Pbr parity suite over the shared test scenes.

• Native Renderer::render_hdr: scene-linear float resolve on the GPU,
  same averaging contract as the scanline backend.

• Resolve averages fall back to the straight mean when every sample is
  fully transparent (scanline parity).

Changed

• GpuRenderer::new / with_backend now request the adapter's
  storage-buffer binding size, buffer size and storage-buffer count
  (up to 8), on top of the downlevel defaults. The upload / draw
  signatures are unchanged.
• GpuRenderer::upload takes &mut self and &RenderOptions (for
  time / animation / material variant).