LASSCF localize_init_guess can't always preserve an ROHF wave function.

[MatthewRHermes]

If the number of active orbitals equals the number of singly-occupied orbitals of a ROHF reference wave function, then with the kwarg freeze_cas_spaces=True the localize_init_guess function will ensure that the ROHF reference wave function exists within the initial guess orbital Hilbert space. However, if doubly-occupied or unoccupied orbitals are included in the active space, then the localization will deoptimize them even if freeze_cas_spaces=True and even if such orbitals are already highly localized (because they will always have at least some tail on atoms outside the fragment).

[MatthewRHermes]

This is not a new issue; rather it is a clarification of the underlying reason why certain seemingly-simple LASSCF orbital optimizations have historically been very difficult.