As per the comments it isn't supported with cache-the-world on, it's
barely documented if at all, and a test for it seems to be failing
without me being able to reproduce it at least (tried on both windows
and Linux).
I don't think it's worth digging too much more into it. My guess is a
reentrant / differently timed a11y event caused by my scheduling
changes, or so, which would override sEventTargetNode or the other
static bool.
The test has ~always failed on macOS too...
Differential Revision: https://phabricator.services.mozilla.com/D210872
e1bbbb144c