Thanks, Martin. Pulling together a summary of discussions like this is very useful.

Regarding Erasmus, I think we need to distinguish two problems.

One is that there may be tagging in EEBO TCP which can
be simplified algorithmically, or hand-improved, so as to let the number of problematic
instances in that corpus decrease hugely.

The other is that, however seldom it occurs, the marginal label 
(I use the word advisedly, as opposed to “note”) with
extended structure definitely exists as a phenomenon which someone,
somewhere, will want to render in HTML.