Sorry to answer my own question here, but in the hopes that this may help somebody else, what I ended up doing was using IHTMLDocument directly rather than IWebBrowser. IWebBrowser is a superset of IHTMLDocument, and the navigation model implemented by IWebBrowser isn't customizable to the degree I wanted.

I actually got MS Developer Support involved and this approach was their recommendation. They say this is what Outlook uses for HTML-based email, which is the user experience I wanted to emulate. They also confirmed that there's no reliable way to filter the OnBeforeNavigate events that result from user action from those that result from script activity.

Hope this helps anybody facing the same issues. It wasn't too hard to port the code to use IHTMLDocument. If you end up doing this, you may also find yourself looking for a way to figure out when the document is done loading. To do that, hook HTMLDocumentEvents instead of DWebBrowserEvents, and look for the DISPID_HTMLDOCUMENTEVENTS_ONREADYSTATECHANGE event. It doesn't tell you what the ready state is; you need to call IHTMLDocument::get_readyState and parse the resulting string. Goofy, but there you go.