?? Who should fix it that it works like expected, i.e. one text node in one p element?
XSLT processors generally hide the issue, whether by normalizing the input while or after the tree is built or by other means.
The interesting stuff of course are XPath processors working on potentially unnormalized DOM structures, in particular Cocoons XInclude processor and the XPath transformer. I'd say you'll get all kind of interesting behaviour.
J.Pietschmann