:arrow_up: I made all TEXCOORD semantics float4, or at least large enough so that no more than one input semantic can fit within the same register.
Now the tests should pass (could you check again @giomasce?), but we still have to address the problem of properly allocating input semantics and see if the shader runner/runtime is replicating the same behavior than native when matching vertex shader output semantics with pixel shader input semantics. I think that's better left for another MR though.