Reflowing papers often produces subpar results. The authors often pay attention to details like on which page certain tables and figures end up, that equations appear where they should etc. I've seen many attempts at rendering HTML from e.g. Arxiv papers and the layout is almost always broken. Paper authors often employ a lot of Latex wizardry to get precisely the layout that they want, and these conversion tools mostly choke on that.
This is definitely true. There are also other things that are very difficult to get to work for math papers in HTML (even with conversions) such complex TikZ and other diagrams. Also, HTML rendering is still pretty ugly compared to LaTeX PDFs because LaTeX has a much better typesetting engine including hyphenation. With subtle changes meaning huge semantic differences in math diagrams, I still much prefer to read PDFs.
Right, changing the equation fonts may also create ambiguity, some custom, complicated equations (several levels of super and subscript, various bars etc) might use some latex hacks or workarounds.
It would require that the authors also create and polish an HTML version, which is extra work.
Even on the official IEEE sites papers look way worse in the HTML view, even though I guess someone at least tries to tweak the to appear okayish.
Purists would argue to separate the "presentation layer" from the underlying content, but it's never like that in the messy real world. Content and presentation are very much intertwined and abstractions are too leaky.
