It isn't niche. I'm an academic writer and there are thousands like me. Like all typst commentators on this site, you have literally just made something up. As for books, you could hardly be more wrong -- /thousands/ of self-published epubs on Amazon are written in absolutely nothing but markdown of one of the specialized varieties. The familiar markdown parsers know how to write out an epub. Self-publishing writers of all sorts have been principal participants of the lists for almost two decades.
I would rather use microsoft word than write in raw typst, which is no different from trying to think in raw latex - even if in some fields it is necessary. I certainly cannot write unless I can output to docx.
A hundred things would be completely impossible to me without a 'lifeless' text format like markdown. The inclusion of a turing complete programming language into document composition - the principal labor distinctive of advanced civilization - is obviously pure insanity, even if it is necessary in the mathematical case. What one needs are ways of introducing the exact features that e.g. wordperfect 5.1 had already completely perfected: paragraphing, sectioning, noting, emphasizing, citing, etc. These are the internal semantic features of actual expression of actual thought, things that must be preserved e.g. by a translator - as sexy typesetting need not be. Considier the question: What is that which must be preserved by a translator? It is the question: What is thought? If I am typing typsetting commands not relevant to this, I am not writing, I am not thinking, I am not constructing a reasoned argument: I am doing high quality page-scale finger painting.
For example, I now consider it intellectual malpractice to give advanced students a translated text for close study, without also supplying the original in parallel. I can do this by simply zipping two markdown files together header to header, paragraph to paragraph with a simple script making html - the original in, say, smaller type narrower column at left margin.
The translation can be changed or an alternative constructed or a new one produced automatically by one of the translation devices. This is an operation that is basically impossible with genuine typesetting, though each md file compiles to pdf e.g. via typst after a markdown->typst process.
This is quite as impossible in typst as it is in latex - huge projects have been devoted to it in latex with complete failure - the necessity of pagination introduces the necessity of precise knowledge of both languages if the page break happens ... in medias sentence - such people e.g. the Loeb library used to have. With html it's a complete picnic, a literal zip operation with a markdown reader -- but of course isn't proper typsetting. The merit is that html is unpaginated. The file for each language itself makes a perfectly sound typst pdf with a keystroke.
How are you doing meaning-aware pagination if you're zipping together markdown columns?
A few minutes of asking an AI and writing a pair of simple typst docs with a few sections of #lorem() paragraphs resulted in what I think your ad-hoc scripting (in what language? add and your custom scripting to your collaborators' publishing stack) efforts do: a typst meta-document that pairs two simpler typst documents, section by section, paragraph by paragraph, each sub-document with its own column width and font size, paginated on paragraph breaks. You might balk at typst's embedded language, but it does avoid added dependencies and external scripting, and the language looks fairly simple a nice and functional.
It might require a somewhat different design to accommodate a marginal-note kind of thing with irregular vertical spacing not paragraph-aligned, but I suspect that's quite possible too, probably with some markup to label notes and align them vertically with corresponding labels in the primary text.