Accessibility in PDFs is also very difficult. I'm not sure any publishers are yet meeting PDF/UA-2 requirements for tagged PDFs, which include things like embedding MathML representations of all mathematics so screenreaders can parse the math. LaTeX only supports this experimentally, and few other tools support it at all.

The reason this doesn't work in practice is that authors don't always play nicely, not because of bad intentions, but because they don't want to cooperate but because of the realities of life: they don't have the time to study style guidelines in detail, they use their own auxiliary LaTeX macro collection because that's what they're used to, or simply because of oversights. Also, typesetting often includes a whole lot of meticulous things, if you listed them all in a guide sheet, that would be a long list of stuff at a level that's too detailed for authors.

I'm not saying it's impossible for authors to fully follow a publisher's style guide but there's a reason publishers employ full time workers who do nothing else but correct submitted manuscripts. Like many other professions, it's a trained skill.

The authors write up their research results.

The editors organize the review process together with the reviewers and the publishing process together with the publisher.

The reviewers read the papers and write their reviews.

The publishers publish the papers.

Stylesheets are typically provided by the publishers and passed on to the authors early on. The reason is two-fold: for one, the publisher wants to produce a high-quality product and uniformity of layouts and styles is an important factor. But the second reason has to do with everything that happens before the publishers even comes into play: common style-sheets also provide some level of fairness because they make the papers by different authors comparable to some degree, e.g., via the max length of a paper.

On top of that, authors often want to present their research in a specific way, and often have strong opinions about e.g. how their formulas are typeset, what aligns with what else, etc. and typically spend quite a bit of time tweaking their documents to look the way they want it. That is, the authors already have an interest in using something more powerful than Markdown.

But like I wrote in another comment here, in doing so, authors do not always adhere to the style guides provided by the publisher - not necessarily maliciously, but the result is the same. For instance, authors might simply be used to handling whitespace a certain way - because that's how they always do it. But if that clashes with the publisher's guidelines, it's one of the things the publisher has to correct in typesetting.

So, perhaps that's the confusion here also to some degree: the typesetting done by a publisher is in the majority of the cases on a very fine-grained level. A lot of is is simply enforcing the rules that were missed by the authors (with the goal of fairness, comparability, and conformity) and small perfectionist's edits that you might not even notice at a casual glance but that typesetters are trained to spot.