I think point #2 is only true for LaTeX-heavy fields like math, physics, and statistics. It's certainly not true for biology, medicine, psychology, or other fields where Word is king.

In biology and medicine, NIH open-access requirements mean most articles are converted to XML (http://jats.nlm.nih.gov/). Articles are submitted in Word files, edited and converted to XML via some proprietary complicated toolchain, typeset to PDF with a big proprietary system like Arbortext, and shipped off to PubMed for archiving. Figures are usually TIFFs or PDFs converted to whatever formats are needed for the web or print.

I don't know how the conversion process works, but I suspect that turning arbitrary Word files into semantic XML with accurate bibliographic cross-referencing metadata and high-quality tables and figures requires a lot of manual work. You have to deal with complicated tables, weird Word formatting, and whatever weird styling the authors think is a good idea. The bibliography has to be annotated with DOIs and converted to a standard format.

LaTeX and BibTeX, or some other structured markup language, are undoubtedly much easier to deal with and much easier to automate.

I think Michael Eisen of PLOS has commented on the situation and said they're actively working on rebuilding their publishing toolchain to automate more of the process, since it requires hours of work for every paper.