There is some correlation in the hash function itself.

Original CMJ hash:

Current simplified version of CMJ hash:

Weirdly enough this was still solving the issue.

Here's some other function i came up with:

Seems much better and it is same number of math operations.

Benchmark results:

BMW: +0.5%
Classroom: +0.6%
Fishy Cat: +0.2%
Koro: +1.7%
Barcelona: -0.4%

For the Koro it's a bit too much, but i've got upcoming AVX2 optimization for hair which will compensate for that.

That seems quite acceptable, especially if solving this correlation helps with convergence in some cases.

I can't judge it well without running the tests, but if that plot tells the whole story then we can entirely replace cmj_hash() with cmj_hash_simple(), which would speed up CMJ sampling and branched path tracing.