As part of the 2015-16 Q2 experimental goal, we did some analysis of HTML content for a sample of articles. The HTML output was mostly driven by Parsoid, however, so we should redo this analysis but with the MediaWiki parser as the backend.
As discussed @Jhernandez I took a look at using MobileFormatter and results are somewhat worrying - https://phabricator.wikimedia.org/T110613#1946285 but trade off for first paint may be worth pursuing...
Analysis published: http://chimeces.com/loot-content-analysis/
No surprises, results are very similar to restbase ones (a bit less payload size on references in general).
Will send email soon to mobile-l and consider how to perform this kind of work on a bigger sample set (doesn't need to be the whole wikipedia, but a much bigger sample set).
Email sent to mobile-l: https://lists.wikimedia.org/pipermail/mobile-l/2016-January/010050.html
This has been in Ready for Signoff since last Wednesday. The actionable AC – all but the last – have been met: we're prioritising stripping navboxes and we're investigating stripping references.