gpu_probe

What GPU a session has here, for whatever brings one up.

A Wayland compositor needs a working GBM/EGL stack on a DRM render node, and device paths cannot answer whether that stack works -- a node can exist with nothing usable behind it, and a driver can be installed with no node at all -- so the answer comes from pixelflux running the compositor's own renderer bring-up. That bring-up also resolves render_dri and the auto_gpu token exactly as the capture will, so the GPU it reports is the one the session uses.

Prints what it found as KEY=VALUE lines on stdout, for a caller to export and act on, and one line saying it in words on stderr:

SELKIES_GPU_RENDER_NODE  the node the session renders on, empty for none
SELKIES_GPU_DRIVER       the driver behind it, empty when no GPU backs it
SELKIES_GPU_PRESENT      whether a GPU is here at all
SELKIES_GPU_ACCELERATED  whether the compositor's renderer came up on it
SELKIES_GPU_GL_VENDOR    the glvnd vendor library whose EGL brings that GPU up
SELKIES_GPU_EGL_X11      the first glvnd vendor whose EGL initializes on an X display
SELKIES_GPU_MESA_DRIVER  how Mesa reaches the GPU: `native`, `zink`, or "" for software
SELKIES_GPU_VULKAN_PRESENTS  whether the GPU's Vulkan driver gave an X window a swapchain, when Zink is Mesa's path

The client-side facts come from bringing each vendor's EGL up the way an application would: on the render node through the device platform, and on a throwaway X server for the window paths -- one drawing on the node through glamor and DRI3 where the session's will -- so a driver that renders but cannot present is reported as such rather than guessed from device files.

What to do about that -- which display backend to start, which GL stack the session's applications are given -- is the deployment's own policy, and belongs where the deployment is assembled rather than here.

Exits non-zero without printing anything when no report can be had: a driver that refuses to answer at all.