As I said in the original thread, the problem is that roundoff errors usually make it impossible to distinguish semidefinite matrices from slightly indefinite matrices. I looked through the CHOLMOD documentation and I don’t see any option to specify a tolerance to ignore slightly negative pivots. Pivoting doesn’t help if it’s indefinite. So the problem is that even with check=false the Cholesky factorization might simply get stuck partway through.
You could try an L D L^T factorization instead. But even if you have a factorization that succeeds, I wouldn’t use the normal equations in an ill-conditioned case: as soon as you form A^T A you have lost too many digits.