The first idea makes sense, but because of how we've already architected things (and the direction we're trying to rearchitect things) it isn't really that feasible.

The second idea could be good. It isn't that hard to make globally named handles that can be reopened in the child process, and that avoids the need for a coherent inheritance chain of processes. Maybe it could break other scenarios that do rely on inheritance though? (But aren't those already broken? All this is *just* outside the edge of my experience, so I'd have to try them all out to be sure.)

It's a regression in 3.7.2, which is when the venv script changed. As I said, updating from 3.7.1 to 3.7.2 was going to change the venv "script" anyway (which was just the main Python executable and its binaries), so it should have been no more breaking than that. But it was, so I consider it a regression (in venv, to be clear, not in multiprocessing).