This should be addressed for the 3.2 release, otherwise the 3.2 release will crash on some machines with Intel graphics. (Blender dereferencing a NULL pointer from a failed shader compilation)

Quick notes:

Increasing the number of transparent render bounces makes CPU and GPU renders match again.

float3 wr = -wi;
followed by
normalize(wi + wr)
inevitably causes a division by zero. In this instance, it means that throughput is (1, 1, 1) at all times, making the call to interpolate_fresnel_color() redundant.

An SBVH, where the same primitive can be in multiple BVH nodes, is a scenario where I can imagine the same triangle being tested twice for intersection.

It's the decorrelation for higher dimensions that I copied from our Sobol sampler that appears to be causing this.