I would note arXiv requires the source as well, and having the source is what is enabling the HTML experiments they're doing.

On consistency, what the journals provide is some level of QA (how much is a function of field and journal, rather than the what is charges), and the template is the journal's brand, so both the authors and journals benefit from the style (I can tell the difference between the different (all similar quality) journals in my field at a glance by the style).

It's also worth noting that there's a whole much of metadata that needs to be collected (whether you agree with it or not, funders require it), so a PDF isn't going to cut it here either.