Idea: Large-assembly performance — instancing for 10K+ parts

[ecto]

LOD generation (packages/engine/src/gpu.ts) + mesh cache exist; no verified scaling test for thousands of instances. Need a large-assembly stress suite + optimisation pass.