[guided]The goal is to convert global $\operatorname{Ad}$-invariance into its infinitesimal form. We do this by moving in $G$ along the one-parameter subgroup determined by $Y$. For each slot, define the smooth curve
\begin{align*}
\gamma_j:\mathbb R\to \mathfrak g
\end{align*}
by
\begin{align*}
\gamma_j(t)=\operatorname{Ad}_{\exp(tY)}X_j.
\end{align*}
The curve starts at $X_j$, because $\exp(0Y)=e$ and $\operatorname{Ad}_e$ is the identity map on $\mathfrak g$. Thus
\begin{align*}
\gamma_j(0)=X_j.
\end{align*}
Its derivative at $0$ is the infinitesimal adjoint action:
\begin{align*}
\gamma_j'(0)=[Y,X_j].
\end{align*}
Now package the invariant expression into a single real-valued function. Define
\begin{align*}
F:\mathbb R\to \mathbb R
\end{align*}
by
\begin{align*}
F(t)=P(\gamma_1(t),\dots,\gamma_k(t)).
\end{align*}
Because the previous step proved that the polarized form $P$ is invariant under the simultaneous adjoint action on all slots, we have
\begin{align*}
F(t)=P(\operatorname{Ad}_{\exp(tY)}X_1,\dots,\operatorname{Ad}_{\exp(tY)}X_k)=P(X_1,\dots,X_k)
\end{align*}
for every $t\in\mathbb R$. Hence $F$ is constant, and therefore
\begin{align*}
F'(0)=0.
\end{align*}
It remains to compute this same derivative using multilinearity. Since $P$ is $k$-linear, the derivative of $P(\gamma_1(t),\dots,\gamma_k(t))$ is obtained by differentiating one slot at a time while holding the other slots fixed. At $t=0$ this gives
\begin{align*}
F'(0)=\sum_{j=1}^k P(\gamma_1(0),\dots,\gamma_{j-1}(0),\gamma_j'(0),\gamma_{j+1}(0),\dots,\gamma_k(0)).
\end{align*}
Substituting the two identities $\gamma_i(0)=X_i$ and $\gamma_j'(0)=[Y,X_j]$ gives
\begin{align*}
F'(0)=\sum_{j=1}^k P(X_1,\dots,X_{j-1},[Y,X_j],X_{j+1},\dots,X_k).
\end{align*}
Since $F'(0)=0$, this is exactly
\begin{align*}
\sum_{j=1}^k P(X_1,\dots,X_{j-1},[Y,X_j],X_{j+1},\dots,X_k)=0.
\end{align*}[/guided]