A content model is the HTML standard's description of what an element may contain. It is stated in terms of content categories: flow content (almost everything in body), phrasing content (text and inline-level elements like span, a, em, img), plus metadata, sectioning, heading, embedded and interactive content. <p>'s content model is phrasing content; <div> is flow content but not phrasing, so <div> inside <p> is non-conforming.
Some elements are transparent: <a>, <ins>, <del> take the content model of their parent. That is why <a href><div>…</div></a> is valid where its parent allows flow content, but <p><a><div> is not.
Content models are conformance rules for authors; validators report them. The browser's parser has its own rules, and it repairs only the cases its tree-construction algorithm names. For <p> the rule is concrete: in the "in body" insertion mode, a start tag such as div, ul, section, table, h1–h6 or another p first checks for an open p "in button scope" and, if there is one, closes the p. That is also why <p>'s end tag is optional. Then the stray </p> finds no open p, which is a parse error, and the parser inserts an empty <p> and closes it.
So <p>Intro<div>Block</div></p> produces three siblings: <p>Intro</p>, <div>Block</div>, <p></p>. CSS like p > div never matches, and the empty paragraph adds its default margins.
Other violations are left alone. <span><div> and <button><div> stay nested exactly as written. The parser also rewrites a few more cases: a second <a> closes the first, a nested <form> start tag is ignored, and non-table content inside <table> is foster-parented to just before the table.