Aardvark.SceneGraph 5.7.0-prerelease0022

• • Aardvark.Base.Incremental (>= 5.3.23 && < 5.4.0)
  • Aardvark.Geometry (>= 5.3.23 && < 5.4.0)
  • Aardvark.Rendering (= 5.7.0-prerelease0022)
  • FShade.Core (>= 5.7.9 && < 5.8.0)
  • FShade.GLSL (>= 5.7.9)
  • FSharp.Core (>= 8.0.300)

- [Sg] Heap: the per-slot derived-uniform / trafo-stack-collapse compute pass gained a **df32** (double-float = two float32) path, so it runs on backends with no shader `double` type (MoltenVK/Metal). `composeModel`/`composeModelInv`/`composeDerived` have df32 variants that read each constituent as a `(hi,lo)` pair from the f32 arena and fold in df32 (Veltkamp-split TwoProduct + fma-guarded sums); the CPU packs constituents and chain links as `(hi,lo)` to match. The path is chosen per bucket from the new `IRuntime.ShaderDouble`: real fp64 (M44d) where the backend has shaderFloat64, df32 where it does not — or forced everywhere via `Heap.ForceDf32` / the `AARDVARK_HEAP_DF32` env var (for desktop validation). The fp64 path is UNCHANGED and remains the default on desktop; with df32 forced on desktop Vulkan the heap golden suite is byte-identical to fp64 (maxDelta=0, sgheap/livechain/sgchain/… all pass). NOTE: df32 is desktop-validated (algorithm + FShade→SPIR-V lowering); validation on an actual Metal/MoltenVK device (fast-math / contraction survival of the error-free transforms) is still pending. A shader that requests a uniform at `double` precision DIRECTLY (in-shader `M44d`/`V3d` read) is unaffected and stays desktop-fp64-only for now.
- [Rendering] Added `IRuntime.ShaderDouble : bool` — whether shaders can use 64-bit floats (Vulkan = `shaderFloat64`; GL = GLSL ≥ 4.0; false on MoltenVK). Lets consumers branch on shader double support (the heap uses it to pick its df32 vs fp64 derive path).