Yes, if you want to stick to Markdown (as opposed to going with the (I believe) stricter RST, Asciidoc or org-mode), then limiting your own markdown via prettification or linting is the way to go, not trying to establish a new Markdown variant.
You mention training a neural net. I’ve also used AI to explain why Markdown is the way it is. I strongly believe this to be the reason Markdown is so popular. Let’s not destroy that, especially since there already exist numerous Markdown-like “strict” code-like formats. In fact I think we can take Markdown’s “pseudo AI” even further down the try to interpret the plain text with human eyes path. I’m even working on it. Though maybe a trained NN would be better than pseudo-AI heuristics.
I beg to differ and most python programmers would take your claim as not just wrong, but insulting. In any case, Markdown is in no way analogous to Python.
The thing is the rules are only complicated for pretty extreme corner cases. Markdown more than meets the 80/20 rule for what it is attempting to do. Probably 95/5 or even 99/1. Most writers have no problem with it. It’s mainly parser programmers who complain 
. I believe the flaws Markdown/CommonMark has (like the rules for lazy continuation) are flawed because they don’t follow the original principle about readability for humans enough, and as I mentioned above I would fix it in the opposite direction you advocate.