Renderer

The renderer sits behind the JARVIG RHI. It does not call DOM, browser, or wgpu types. The first GPU backend is wgpu, and native/jarvig_rhi_wgpu is the only crate where those types may appear. D3D12, Vulkan, and Metal are what that backend targets, not extra public backends, until a platform forces a direct one. See rhi.md and ADR-0019.

The presented path is two depth-tested meshes (JRV-0047) with engine-owned GPU lifetimes (JRV-0048), drawn by two views of one world (JRV-0049). The world is extracted once into a snapshot (JRV-0050). The renderer does not query that world while it draws. Surface color comes from a material graph compiled to IR and then to one shared pipeline (JRV-0051, JRV-0052). JRV-0053 draws that through one StandardMetalRough master. JRV-0054 lights it from world lights extracted once and converted per view. See lighting.md, ../materials/pbr.md, and ADR-0029. A mesh does not own that material or those textures. Color space is texture metadata. ADR-0027. A mesh is local geometry plus submeshes and local bounds. It is not an asset, an entity, or a GPU buffer. The logical mesh has no wgpu type and no RHI type. Vertices are object-local meters. Each view turns that into its own camera-relative float32. Perspective depth is reversed-Z with no finite far plane. A view is not a world and not a swapchain. See views.md. The snapshot is not a second world. See scene-extraction.md. The ladder through the standard surface, the first direct lights, and graphics-adapter policy is recorded in the backlog. That numbered renderer ladder stops at JRV-0055, which is accepted. The product milestone is the native editor, JRV-0057. The first shell is JRV-0058. Adapter policy is ../platform/graphics-device-selection.md. Adapter policy is not part of the editor tickets. GPU-driven culling and virtual geometry stay behind it. The residency model they will share is resources.md.

WebGL2 can remain a compatibility backend later. It must not force the CPU to submit tens of thousands of individual draws. ADR-0006 still holds as "do not design around one draw per object." Its rejection of a native graphics API is superseded.

World extraction
  -> GPU scene database
  -> frustum / layer filtering
  -> hierarchical-Z occlusion
  -> instance and cluster selection (compute)
  -> streaming feedback
  -> indirect command generation
  -> depth / visibility / GBuffer
  -> lighting / shadows / GI
  -> transparency / VFX
  -> post
  -> UI / debug

Package @jarvig/render registers module id render, requires world, and counts prepare/execute hooks. The default device is { backend: 'null', label: 'null-device' }. Server profiles never call those hooks, and the server package does not import this module.

Promotion sequence

Do not skip ahead.

1. Correct PBR baseline and render graph.
2. GPU scene and indirect rendering.
3. Hierarchical-Z and compute visibility.
4. Meshlet / cluster hierarchy.
5. Virtual geometry paging.
6. Virtual textures.
7. Advanced GI and reflections.
8. Experimental representation virtualization.

Each experimental step needs benchmarks: frame time, GPU time, CPU time, memory, VRAM, streaming bandwidth, storage size, shader cost, visual error, pop-in, and generation cost. Never fabricate them.

Research flags

virtualGeometry, proceduralMicrogeometry, representationVirtualization, and advancedGi default off in the TypeScript prototype. ADR-0061 is the native editor runtime: meshlet drawing, the cluster hierarchy, frustum culling, and hierarchical-Z start on, and microgeometry mode defaults to Auto. Pages: webgpu.md, gpu-scene.md, visibility.md, meshlets.md, virtual-geometry.md, microgeometry.md, representation-virtualization.md, lighting.md.