The method attribute pspace_size of FCI objects indicates the maximum CI vector size for which the Hamiltonian will be diagonalized non-iteratively. At one point, it was obviously supposed to also indicate the size of the lower-energy diagonal block in the "pspace preconditioner", as described by CPL 169 463. However, the pspace preconditioner is both disabled and broken. Disabled, because here:

the pspace Hamiltonian information is deleted and not passed on to the make_precond, so we always branch to the "diagonal" preconditioner. I believe this was broken around the time that support for high-symmetry FCI became a priority, and it makes sense: applying cylindrical or spherical symmetry to the pspace Hamiltonian sounds like a bear. But separately from that, it's also broken, because in this code:

the line that is supposed to actually multiply the inverted p-block Hamiltonian matrix into the residual vector is commented out. The only contribution in this function of the p-block Hamiltonian matrix is to the scalar e1, which just scales the component of the new vector which is proportional to the old vector - a component that is supposed to be zero anyway. The meaningful component of the output x1 vector is identical to the result from the diagonal preconditioner.

This broken code is unused in the PySCF core modules, but the csf_fci module in PySCF-Forge makes reference to it. So, question. If I want to have a better preconditioner in csf_fci, should I:

1. Copy, paste, and fix the unused broken code from the PySCF core modules into PySCF-Forge, even though copy-paste-fix is a dangerous habit to get into, or
2. Post a PR "fixing" this unused code to PySCF core, even though absolutely nothing in PySCF core uses it and I do not know how I would validate the accuracy of the "fix"?