at-spi's modifier relatch behavior is overzealous when GOK is running in
corepointer mode; because the consumed buttonpress triggers an XKB
'relatch'.  That's desirable, but because at-spi can't actually successfull
do an XGrabButton anymore (metacity, I think), it has to use fallback
detection methodsfor button state changes.  It catches the button press,
sends it to GOK which consumes it, then orders a relatch (since it knows
that the button event will trigger an Xkb delatch event as a side-effect,
which we want to revert.

However, in the case of GOK, the button press also triggers a (synchronous)
call to SPI_generateKeystroke, which in most cases will also
(appropriately!) cause an XKB delatch.  But since we've just issued a
pending 'relatch' request, this delatch will get reverted.

The only solution seems to be to check in generateKeystroke to see if the
generated key event should cause a delatch or not, and "cancel" the relatch
order if it should.  Ugh!