[guided]The goal is to show that every root of $P_{\alpha, F}$ belongs to $L$. The strategy is to leverage the fact that $P_{\alpha, K}$ already splits over $L$ (by normality of $L/K$) and that $P_{\alpha, F}$ divides $P_{\alpha, K}$.
Since $L/K$ is normal, for any $\alpha \in L$, the minimal polynomial $P_{\alpha, K} \in K[t]$ splits completely over $L$. This means we can write
\begin{align*}
P_{\alpha, K} = (t - \beta_1)(t - \beta_2) \cdots (t - \beta_d) \quad \text{in } L[t],
\end{align*}
where $d = \deg P_{\alpha, K}$ and every $\beta_i$ belongs to $L$.
Now, we established in the previous step that $P_{\alpha, F} \mid P_{\alpha, K}$ in $F[t]$, and hence also in $L[t]$ (since $F \subset L$). The polynomial ring $L[t]$ is a UFD (in fact a PID), and the complete factorisation of $P_{\alpha, K}$ into irreducibles in $L[t]$ is the product of linear factors $(t - \beta_1) \cdots (t - \beta_d)$. Any monic divisor of $P_{\alpha, K}$ in $L[t]$ must therefore be a product of some subset of these linear factors. In particular, $P_{\alpha, F}$ is a product of linear factors $(t - \beta_{i_j})$ with each $\beta_{i_j} \in L$.
This means every root of $P_{\alpha, F}$ lies in $L$, so $P_{\alpha, F}$ splits completely over $L$.
Why is the divisibility relation so powerful here? Without it, we would need to argue directly that the roots of $P_{\alpha, F}$ lie in $L$ — but $P_{\alpha, F}$ is defined over $F$, not over $K$, so the normality of $L/K$ does not apply to it directly. The divisibility $P_{\alpha, F} \mid P_{\alpha, K}$ bridges this gap: it ensures that the roots of $P_{\alpha, F}$ are a subset of the roots of $P_{\alpha, K}$, and normality of $L/K$ guarantees that all roots of $P_{\alpha, K}$ lie in $L$.[/guided]