bug#81564: Possible bug when binding dynamically with an existing lexical binding

The code

  (let ((x 2)) (defvar x) (let ((x 10)) x))

returns 2, which seems extremely unintuitive to me, if not an outright
bug.

The unintuitive thing is that the `defvar' effects not the first binding
but the second binding, but if that is simply how it works then I don't
have anything to say about it.

The possible bug is that the `x' that returns is referring to the
lexical `x' bound in the outer `let', instead of referring to the
dynamic `x' bound in the inner `let'.

I think what is happening is that when `x' reaches `eval_sub', it looks
for the lexical value first, assuming that it wouldn't have a lexical
value if it was marked dynamic, not taking into account the situation
where it was marked dynamic only locally with a lexical binding around.

I am unsure if this is a bug or simply an illegal way to use `defvar'.