Come to think of it, perhaps the problem is more that we equate an "XML document" to a "web resource". What we perhaps need is a way of distributing a single XML document over a large collection of web resources, and then navigating around that XML document seamlessly, using XPath?

Of course we can do that crudely already, using entities or XInclude. Perhaps we just need a smarter implementation of transclusion, where the document fragments are fetched on demand when XPath navigation needs them, rather than being assembled eagerly by the XML parser.