Casting (some) links
To the extent that arbitrary cross-references and other links are represented in WordML, XSweet captures them and render them as HTML
<a> may also be generated by XSweet because it "needs to do so", for example to "glue together" a footnote with its referencing string.
There are two categories of potential links in the Word source, which fall into neither category. They could be marked with inline formatting (say, maybe blue and underlined?) which is expected to be "seen" as a link. In such cases there will be some indication that this is a link, but it won't be known ahead of time -- or necessarily reliable (bad data). Or worse (second category) - the link could be present but not explicitly marked in any way -- just an "http:// ..." run in otherwise plain text.
The first case is an "arbitrary mapping " problem and goes back to #3 (closed).
We could address the second case in a filter that detected and called out target substrings with a regular expression. Due to complications of mixed content, this is a bit trickier to define than a generic element/element mapping, so it has to be done with some care (probably in a separate module), but it is technically feasible. (This is the machine equivalent of a regex-based search/replace -- it works as long as there is no conflicting markup already there. BTW we can do this of course for any sort of matched content, not just links.)
(Note that this requirement is most common when dealing with citations, which puts it into an entirely different context from the 'main body' of text import.)