> Changing the read technique mid-transform would be a bit strange, even
> if it was conformant

It wouldn't be conformant.
document order of attribute nodes is stable and well defined, given an
input tree. What isn't defined is how the nodes are ordered in an input
tree, given the XML document markup. So once teh document has been
parsed, attribute order is just as fixed as eleemnt order.

are semantically equivalent but the result of string-join(@*,'-')
would (I think, most of the time) be different, so it really depends
on the consistency of the ordering in the on-disk version, and the
order in which the XML parser reports the attributes.