CURSOR_DECLARED was set in capture_virtual_output, which is not on the Windows
prep path, so it was never true and the between-session clean returned at its
first gate every time, silently. Both on-glass runs that appeared to 'not fire'
were this; the one run that did fire was a build with no gate at all.
Move it to the ADD request in the driver module, which every session that wants
a hardware cursor goes through by construction.