> I guess we could try and be smarter about the error recovery, specifically,
by putting a filter in front of the complex content constructor to modify
invalid content (like attribute nodes) before they get in there and trigger
the error.
>

In fact there already is such a filter - it attempts to turn the attribute
nodes into processing instructions - but the filter is misplaced in the
processing pipeline, and this error is triggered before the filter gets in to
intercept it. Will fix.

Michael Kay
Saxonica