GContextSpecificGroup has been somewhat broken for a rather long time:
when we remove the last reference on an object held in the group, we try
to clean up the source, but fail to actually remove it from the
mainloop.

We will soon stop emitting signals on the source (due to it having been
removed from the hash table) but any "in flight" signals will still be
delivered on the source, which continues to exist.  This is a problem if
the event is being delivered just as the object is being destroyed.

This also means that we leave the source attached to the mainloop
forever (and next time will create a new one)...

This is demonstrated with the GtkAppChooser dialog which writes an
update to the mimeapps.list file just as it is closing, triggering the
app info monitor to fire just as it is being destroyed.

Karl Tomlinson correctly analysed the problem and proposed this fix.

https://bugzilla.gnome.org/show_bug.cgi?id=762994