thank you for the explanation. However, I still suspect that the observed behaviour is a bug and in contracdiction to the 2007 spec.
The quoted text is not the text of the erratum XT.E37, but the text of the 2007 version (which is identical to the previous 2006 version). To make it clear: the text of the erratum would be the explanation of the observed behaviour - it restricts the namespace copying to named nodes:
XT.E37: "... then each named (that is, non-default) namespace node ... is copied"
And this exception is exactly what Saxon does - the xmlns="" shows that the nameless namespace node has *not* been copied.
The 2007 text, however, does not make this distinction, and therefore the namespace *must* be copied and the xmlns="" is in contradiction to that.
PS: Although it might seem so, this is not at all an academic issue: some schema editing solutions have to rely on a complete construction of the original namespace bindings in newly constructed elements.
PPS: As a psychological sidenote, I find it remarkable that the bugzilla text (which of course I did not know) uses an attribute "abcd" in order to demonstrate the problem, whereas I invented an attribute "abc", funny, isn't it. Du gleichst dem Geist den du begreifst, Goethe said.