Order of our service teardown is somewhat undefined. kglobalaccel5 can be closed before plasmashell.
On plasma exit we call setInactive/unregister on various kglobalaccel actions and accidentally spawn kglobalaccel again.
If kglobalaccel is already closed, our actions aren't registered. There is no point waking it up to try and remove an action. Use of the setAutoStartService flag was chosen over checking if the kglobalaccel service is still registered as that is still racey.
This was always slightly broken, but changes in the ecosystem means the impact this has has worsened.