The Jordan-Chevalley decompositions holds for semisimple Lie algebras. The Lie algebra $\mathfrak{gl}_n(\Bbb C)$ is not semisimple, since it has a non-trivial abelian ideal, its center.
Furthermore, $d$ need not be diagonal. It is only diagonalizable. And indeed, if $x$ is a non-nilpotent endomorphism in a linear Lie algebra, $ad(x)$ can be nilpotent. For $x=I$ in $\mathfrak{gl}_n(\Bbb C)$, we have $ad(I)=0$, the zero endomorphism, because $[I,B]=IB-BI=0$ for all $B\in \mathfrak{gl}_n(\Bbb C)$.
An element $x\in L$ is called ad-nilpotent, if $ad(x)$ is nilpotent.
"Should I understand this as saying that $ad(x)$ is nilpotent iff $x$ is nilpotent?" No, this is not true, as you have seen yourself.
Took some work and some scouring through the literature, but we got there in the end. Both questions are answered positively — though my brain still needs some time to digest the answer for question one, and make sure that it's really really true, so, eh, approach with care.
First question:
Whenever I say parabolic here, I mean not the Knapp definition, but the one in terms of complexification, see the OP.
From Lemma in Section 3.2 from Wolf, Koranyi, we can extract the following (heavily paraphrased, but hopefully equivalent):
Let $\mathfrak{g}$ be a real semisimple Lie algebra and $\mathfrak{q}$ a parabolic subalgebra. Then there is some Cartan decomposition
$$\mathfrak{g} = \mathfrak{t} \oplus \mathfrak{p} $$
some maximally noncompact ("maximally split") Cartan subalgebra $\mathfrak{h} = \mathfrak{t} \oplus \mathfrak{a}$ of $\mathfrak{g}$, where
$$\mathfrak{t} \subset \mathfrak{k} \text{ (the "totally nonsplit" part)}, \quad \mathfrak{a} \subset \mathfrak{p} \text{ (the "totally split" part) },$$
a subspace $\mathfrak{a}' \subset \mathfrak{a}$ and a choice of positive roots $P$ in the restricted root space decomposition of $(\mathfrak{g}, \mathfrak{a}')$ so that
$$\mathfrak{q} = \mathfrak{g}_0 \oplus \bigoplus_{\alpha \in P} \mathfrak{g}_\alpha, \quad
\mathfrak{g}_\alpha = \{x \in \mathfrak{g} : [a,x] = \alpha(a) \cdot x \quad \forall a \in \mathfrak{a}'\}.
$$
(Careful: In the source, the root space decomposition is carried out in the complexification $\mathfrak{g}_\mathbb{C}$, but we can also carry it out in the real setting, since the ad-action of elements in $\mathfrak{a}' \subset \mathfrak{a}$ is real diagonalizable. This is always necessary for the restricted root space decomposition.)
Very verbose, but in the end, in the above notation, every parabolic subalgebra contains $Z_\mathfrak{k}(\mathfrak{a}) \oplus \mathfrak{a}$ in the $\mathfrak{g}_0$-component and some choice of positive restricted roots of $\mathfrak{a}$ in the $\bigoplus_{\alpha \in P} \mathfrak{g}_\alpha$-component. Hence every parabolic subalgebra contains some subalgebra of the form $\mathfrak{m} \oplus \mathfrak{a} \oplus \mathfrak{n}$. And indeed, the subalgebras $\mathfrak{m} \oplus \mathfrak{a} \oplus \mathfrak{n}$ are parabolic, since their complexification contains a Borel algebra associated to the complexification of the Cartan subalgebra $\mathfrak{a} \oplus \mathfrak{t}$.
Hence, Knapp's minimal parabolic subalgebras are exactly the minimal parabolic subalgebras in the usual sense.
Second question:
In Bourbaki, Chapter VIII, Exercise 3a for §5, we learn:
If $\mathfrak{q}, \mathfrak{p}$ are parabolic subalgebras of a semisimple (real or complex) Lie algebra $\mathfrak{g}$, and $\mathfrak{q} \subset \mathfrak{p}$, then the radical of $\mathfrak{p}$ is contained in the radical of $\mathfrak{q}$.
And in Bourbaki, Chapter VIII, §10, Corollary 2, we learn:
Every subalgebra $\mathfrak{n}$ of a (real or complex) Lie algebra $\mathfrak{g}$, consisting only of nilpotent elements of $\mathfrak{g}$, is contained in the nilradical of a parabolic subalgebra $\mathfrak{q}$.
As a corollary of the two: Given a subalgebra $\mathfrak{n} \subset \mathfrak{g}$, contained in the radical of some parabolic subalgebra $\mathfrak{q}$. But then there is some minimal parabolic $\mathfrak{q}_0 \subset \mathfrak{q}$ with $\mathfrak{n} \subset \text{rad}(\mathfrak{q}) \subset \text{rad}(\mathfrak{q}_0)$.
Wolf, J. A.; Koranyi, A., Generalized Cayley transformation of bounded symmetric domains, Am. J. Math. 87, 899-939 (1965). ZBL0137.27403.
Bourbaki, Nicolas, Elements of mathematics. Lie groups and Lie algebras. Chapters 7 and 8, Berlin: Springer (ISBN 3-540-33939-6). 271 p. (2006). ZBL1181.17001.
Best Answer
Suppose that $g$ is nilpotent, let $h$ be a 2-dimensional subalgebra of $g$, for every $x\in h$, let $ad_x:h\rightarrow h$ defined by $ad_x(y)=[x,y]$. Since $h$ is defined over an algebraic closed field, $ad_x$ has an eigenvalue. $[x,y]=cy$. If $c\neq 0$, it implies for each $n(ad_x)^n(y)\neq 0$. This is impossible since $g$ is nilpotent. Suppose that $h=Vect(x,y)$, $[x,y]=cx+dy$, $(ad_x)^2=[x,cx+dy]=d(cx+dy)$ you deduce recursively that $(ad_x)^{n+1}(y)=d^n(cx+dy)$ since $g$ is nilpotent, $d=0$. This implies that $ad_y(x)=-cx$. Since $ad_y$ is nilpotent, $c=0$. Thus $h$ is commutative.
On the other side, suppose that every subalgebra of dimension 2 is commutative, let $y$ be an eigenvector of $ad_x$, $Vect(x,y)$ is a 2-dimensional subalgebra, thus it is commutative, thus $[x,y]=0$. This implies that $ad_x$ is nilpotent an the theorem of Engel implies the result.