Thanks so much for all the discussion so far.
I'm not so sure about infinite scroll I need to think more about this. It seems wrong to me as there are always a finite amount of sections (so it's never truly infinite or close to looking like infinite) and these sections are already in the HTML so it seems wrong / unnecessary to hide them.
That said I think there are benefits to both approaches of the existing behaviour - one where they are all open (and can be collapsed) and one where they are all collapsed (and can be open)
I think a sticky preference would be best that uses a combination of localStorage and user preferences (the latter taking preference). I think such a setting could be surfaced as a simple toggle control at the bottom of the footer (although I'm not sure what the icon would look like).
We could also imagine 'learning' a preference based on behaviour by a user (do they always open all the sections they come across?)
Personally the current setup only makes sense to me if the page loads quicker due to not serving the html inside sections and loading the content of those sections only when the section is toggled open (ie. lazy loading content of sections). In the current form we serve all the content and due to this there is an inevitable flash of the section collapsing as the JavaScript and entire page has loaded.
I'm not sure I agree with Steven's assessment that this will make navigating between sections difficult - behaviour gets reverted - you close the section to see the next section. This is akin to flicking through a book and flicking to the next page (closing the section) if the heading at the top of the page doesn't interest you. It just means you don't see all the headings in one go which could be a good or bad thing.
Is there an A/B test we could do here? In situation A we show all sections open by default on say the Barack Obama article and in case B show all sections closed by default (note this is a simple line of JavaScript). If we were to do this what would we be optimising for?
* Would it be how many sections are collapsed?
* What % of the article is read (could equate to how far down the article a user gets)?
I think this matter can be solved by collecting data...