Dear maintainer, I have committed a fix upstream. Ryan's suggestion was almost correct; the trick is that the RNG really ought to be seeded exactly once, and in our case has to be else it will break the print feature that generates multiple puzzles per second... and this is threaded code, so a mutex is required to implement that. The best way to do that is to use std::call_once, but I also wrote an alternate fix using GMutex that would be easier to backport for Debian:
https://git.gnome.org/browse/gnome-sudoku/commit/?h=gnome-3-20&id=30065efc87d381991f51c2a2d39b0ffb0b04425b Of course it'd be even better for qqwing to use a modern C++ RNG that doesn't depend on shared state, but this will do for us....
