This is triggering a core parse, where one shouldn't really be necessary. If it requests the Parsoid HTML, this won't cause extra parsing (since it's cached and we'll eventually be saving the Parsoid version anyway).

We could either:

1. Require Parsoid for this, or:
2. Support both paths, but use Parsoid in production.

We already have the logic in ExtLinkExtractor. Bonus points if the extractor does not run twice.