Kill the 1 to 2 second app launch delay caused by waking the dGPU
The mechanism (the GLVND/Vulkan loaders iterate every ICD in /usr/share/vulkan/icd.d and /usr/share/glvnd/egl_vendor.d and wake a sleeping dGPU on the way) is real and is the same trick Arch's External_GPU page uses, and the Intel path is right. I confirmed vulkan-intel currently ships /usr/share/vulkan/icd.d/intel_icd.json. But the AMD filename given is stale: vulkan-radeon ships radeon_icd.json, not radeon_icd.x86_64.json (Mesa dropped the arch suffix), so an AMD-iGPU reader copy-pastes a path that does not exist and loses Vulkan entirely with no error. Also missing: pinning one 64-bit ICD removes Vulkan from 32-bit Steam/Proton, and environment.d only applies to the systemd user session (not TTY logins or sudo).
With these set globally, Vulkan and EGL apps will not see the NVIDIA GPU at all. Steam games and CUDA/ML workloads launched without an explicit override will silently run on the iGPU or fail to find a device.
Symptom
On a hybrid laptop with working RTD3 power management, GUI apps take an extra second or two to appear every time, especially the first launch after idle. Nothing is CPU-bound. The delay is just dead time before the window shows.
Cause
OpenGL/EGL and Vulkan enumerate every candidate device listed in /usr/share/glvnd/egl_vendor.d/ and /usr/share/vulkan/icd.d/. Even when the iGPU config sorts first, the loader still iterates the NVIDIA entry, which wakes the sleeping dGPU (~1 s) and burns battery, before falling back to the iGPU. It is an NVIDIA driver behaviour, not a compositor bug.
Fix
Check what you actually have before pinning anything, because a nonexistent path in VK_DRIVER_FILES silently disables Vulkan:
ls /usr/share/vulkan/icd.d/ /usr/share/glvnd/egl_vendor.d/
Current names: Intel intel_icd.json, AMD radeon_icd.json, NVIDIA nvidia_icd.json (32-bit variants appear here too if you have the lib32 packages).
sudo tee /etc/environment.d/50-mesa-default.conf >/dev/null <<'EOF'
__EGL_VENDOR_LIBRARY_FILENAMES=/usr/share/glvnd/egl_vendor.d/50_mesa.json
__GLX_VENDOR_LIBRARY_NAME=mesa
EOF
That alone removes most of the wake-ups. Only add a Vulkan pin if you still see the delay, and list every ICD you need, e.g. on an Intel iGPU with Steam/Proton:
# append to the same file, using the exact filenames from the ls above
VK_DRIVER_FILES=/usr/share/vulkan/icd.d/intel_icd.json
Alternative that does not break 32-bit: install vulkan-mesa-implicit-layers and use MESA_VK_DEVICE_SELECT=<vendorID>:<deviceID> per app instead.
Log out and back in (/etc/environment.d is read by the systemd user manager, so it does not affect bare TTY logins). Override per-command when you want the dGPU. See the PRIME record.
Verify
Apps open without the extra pause, and cat /sys/bus/pci/devices/0000:01:00.0/power/runtime_status stays suspended while you launch them.