1/2P is probably optimal. Building the entire molecule then conditionally projecting might also be possible. 1P can still go lower as well but imagine optimizing locally not around the global minimum.