Ok, code I started before the gaming, was to make the terminal query KMS to find out the active GPU. Find a GPU with connectors attached to it (means its plugged into a monitor). Use that GPU. Sounds simple enough. And sure enough, it works. Works really well. There were some wrinkles tho. Officially, according to kernel.org, applications open `/dev/gpu/card0` if they are going to do KMS things. Setting modes. Compositors do this. Apps not doing compositing should open `/dev/gpu/render0`, and do their buffer things there. Thing is, kernel.org is full of shit. Maybe that was true in the past, but nowadays, you can just use the primary node `/dev/gpu/card0` in either case. In fact trying to get the render nodes behaving has been a major pain point.