one middle ground might be to over-resolve slightly, and rather than actually compute the precise reachable constraints from pytest targets, just compute all interpreter constraints in the repo.