The pre-baked box_sample_downsample.ps.hlsl declared only
SV_Position as its pixel-shader input. With no preceding TEXCOORD0,
fxc packed SV_Position at input register 0. The fullscreen-quad
vertex shader these PSOs bind emits TEXCOORD0 at output register 0
and SV_Position at output register 1, so on D3D12 the PS input
register 0 had no matching VS output carrying the POSITION system
value. D3D12 links VS-output to PS-input by register at
CreateGraphicsPipelineState time, so the box PSO failed to compile
and the build returned null. UpdateDisplay then bound that null
pipeline (SetPipelineState dereferences its argument), faulting
inside D3D12Core when downsampling was switched on at runtime.
Declare an (unused) TEXCOORD0 input ahead of SV_Position so fxc
packs SV_Position at register 1, matching the vertex shader output
and the display pixel shaders (which likewise keep an unused
TEXCOORD0 at register 0). fxc preserves declared-but-unread inputs
in the signature, so the resolved layout is TEXCOORD0->reg0,
SV_Position->reg1 for every scale variant. This is also what
GenerateBoxSampleDownsampleFragmentShader already emits via
DeclareFragmentEntryPoint(ss, 0, 1, ...). D3D11 was unaffected
because CreatePixelShader does not validate against the bound VS.
Regenerated all 15 box DXBC variants (s2..s16).