Aardvark.SceneGraph.Browser 5.7.0-prerelease0023

• • Aardvark.Application (= 5.7.0-prerelease0023)
  • Aardvark.Base.Tensors (>= 5.3.23 && < 5.4.0)
  • Aardvark.Rendering (= 5.7.0-prerelease0023)
  • Aardvark.Rendering.GL (= 5.7.0-prerelease0023)
  • Aardvark.Rendering.Vulkan (= 5.7.0-prerelease0023)
  • Aardvark.SceneGraph (= 5.7.0-prerelease0023)
  • FShade (>= 5.7.11 && < 5.8.0)
  • FSharp.Core (>= 8.0.300)
  • Offler (>= 2.1.1 && < 2.2.0)

- [Sg] Heap: `ofRenderObjects` allocates nothing until first use and frees everything (GPU + per-object CPU) when the last render task drops it, rebuilding cleanly on reuse — resource lifetime is ref-counted across tasks via a new backend-ignored `ActivationRenderObject` (carries only activate/deactivate). Fixes faults when a heap is reused by a task after a previous one was disposed.
- [Vulkan] ComputeTask: on portability-subset devices (MoltenVK/Metal) replay compute command streams INLINE into the primary command buffer instead of recording them into a reusable secondary and executing it via `vkCmdExecuteCommands` — which SIGSEGVs in MoltenVK (`MVKCmdExecuteCommands::setContent`). Mirrors the existing `CommandTask` `useInline` render-pass workaround (forceable everywhere via `AARDVARK_INLINE_RENDERPASS=1`). Fixes a hard crash when a heap (or any compute-using) scene renders on Apple Silicon. Non-portability backends are unchanged (still record + execute a secondary).
- [Rendering] Transparency wrapper: detect `VK_EXT_fragment_shader_interlock` / `GL_ARB_fragment_shader_interlock` at runtime and pick the OIT technique accordingly. The exact A-buffer path (interlocked per-pixel k-buffer) is used only where interlock is available; everywhere else the wrapper silently falls back to Weighted-Blended OIT. Previously the A-buffer path tried to run unconditionally — on AMD's Windows proprietary driver and most mobile / iGPU drivers (where interlock is absent) the result was a broken or crashing OIT pass. Set `AARDVARK_OIT=wb` to force Weighted-Blended on a backend that does support interlock (useful for A/B comparison or where the A-buffer path is known slow); `AARDVARK_OIT=abuffer` forces A-buffer where supported.
- [Rendering] Added `IRuntime.FragmentShaderInterlock : bool` — exposes whether ordered per-pixel critical sections are available to shaders. Used by the transparency wrapper to choose its OIT technique; useful for any other consumer that needs the same capability.
- [Rendering] A-buffer slot storage switched from a wide `W·Capacity × H` 2D image to a `W × H × Capacity` 2D-ARRAY image (FShade `UIntImage2dArray<rgba32ui>`). The flat layout hit AMD's `maxImageDimension2D` cap (16384) at window width ≥ 2048, causing outright image-creation failure (`Cannot create R32g32b32a32Uint image with size [20480, 1600, 1] (maximum is [16384, 16384, 1])`) and crashing the showcase on the AMD 890M Windows driver and the 9060XT. The 2D-array layout lifts the cap to per-axis (16k on AMD), matches the per-slot semantics naturally (one layer per fragment slot), and produces simpler indexing in the insert/resolve shaders.