This really feels like it should be spamassassin's job.  HTML pages are more likely to trigger spamassassin's html triggers (obviously), and I don't know what it would think about complex pages likely to to be referenced in non-spam emails (news sites, etc).  I'm worried about the false-positive rate, and what effect pulling in all those html pages would have on the bayes database (for both spam and ham).

This patch to the latest spamass-milter helps to stop message which consist of just a URL and maybe a few other messages. When spamass-milter is started with the -U option the URLs contained in the message are scanned for their likely spammism. There are a couple of TODOs and I'm not sure that the autoconf system works with my change (it needs to be CURL aware), but it's good enough for playing with.