This caused 19 dirty diffs on the page https://fr.wikipedia.org/wiki/Wikipédia:Le_Bistro/20_février_2021, listing them here so that the task can be cross-referenced from dtcheck:

180107829 180109981 180110580 180115820 180121164 180129746 180129766 180129784 180129800 180133852 180134443 180134523 180134723 180135513 180137301 180146003 180167226 180167416 180178038

HTML5 tree builder actually allows <div> inside <a> – but it does not allow <div> inside <p>:

This feels very similar to the link-inside-link scenario (T150196), and I know that Parsoid is able to roundtrip that one (although I'm not sure how it does it). Is the trick you use for that not applicable here?

Subbu wrote in IRC:

[[Bar|<div>x</div>]] is the test case.

It is tricky though becuase ... as bartosz pointed out, the problem is not div-in-link ... it is the p-wrapper around the link that cause the parapraph to split before the div.
so, that syntax is valid in a table cell, for ex.
or valid inside a <div>

Given that we have run into this for the very first time after so many years, I am marking it low as being an edge case. But we'll fix it.

In T275415#6849733, @matmarex wrote:

This feels very similar to the link-inside-link scenario (T150196), and I know that Parsoid is able to roundtrip that one (although I'm not sure how it does it). Is the trick you use for that not applicable here?

That works because Parsoid explicitly handles it via this complex logic. The bug is an instance of T165098: unpackDOMFragments doesn't handle all mandatory content model constraints (those that are enforced by the HTML5 tree builder on a HTML5 parse) into which I am going to merge this task next. :-)

Looks like this was the 4th instance of this bug actually! So, maybe we ought to consider fixing this one of these days.