The answer is that the DOM inherently has state. It could be a textbox's text input and its cursor state, or the focus state, or whatever. But developers don't want to think about this DOM-managed state to simplify their mental model.
Because walking the tree sucks, and more specifically building upon the tree sucks a lot (createElement, appendChild, etc). That's why innerHTML, which was _not_ a standard, became a standard after being widely implemented and used.
So, the solution was to almost never read the actual tree, because it is slow and weird. This was solved before React (libs like Backbone and others kept track of state).
Regarding DOM mutation, the browser goes through possibly a lot of stuff (reflow, repaint, etc) when the DOM is changed. React is designed to allow components to optimize this lifecycle. It is very easy to misuse though. You have to know the lifecycle to be able to use it effectively, so, you have to think about the state (or just be allowed to use off-the-shelf components, or be ready for unexpected pitfalls).
It seems something like VDOM could be introduced to the standard DOM API. Some kind of detached DOMDocument that is very sterile and fast. It should be faster than doing it with JS. Remember when we parsed JSON with libs written in JS? Or when CSS selectors were parsed inside jQuery? Do you notice my point?
No, when was that? Because it was always possible to `eval(jsonString)` and get the parsed result. Indeed, that was the whole point.
Would normalize.css exist if the standards were more specific around default styles?
Would jQuery/sizzle exist if CSS selectors were available as a DOM API in the first place?
Would vdom exist if DOM was faster?