[guided]The second fundamental form must be computed from a section of the moving Hodge filtration, not from an arbitrary flat family. We therefore choose a local holomorphic section $\sigma$ of $F^p\mathcal H^k$ whose value at $0$ represents the given class $[\alpha]\in F^p_0/F^{p+1}_0\cong H^{p,q}(X)$. Using the smooth trivialisation, represent $\sigma(t)$ by closed complex $k$-forms
\begin{align*}
\eta_t\in A^k(X;\mathbb C)
\end{align*}
on the fixed smooth manifold $X$. At $t=0$, the class of the fixed $J_0$-type $(p,q)$ component of $\eta_0$ is represented by the chosen Dolbeault-closed form $\alpha\in A^{p,q}(X)$.
Because $\sigma(t)$ lies in $F^p_t$, the form $\eta_t$ has no $J_t$-type component with holomorphic degree less than $p$. The first possible component below $F^p_0$ after expanding in the fixed $J_0$ decomposition is therefore the component of type $(p-1,q+1)$. The Beltrami differential
\begin{align*}
\mu\in A^{0,1}(X,T_X)
\end{align*}
encodes precisely this first-order comparison of type decompositions. For a smooth $k$-form $\omega$ on $X$, let $\operatorname{pr}^{a,b}_{J_0}\omega$ denote the projection of its fixed $J_0$-type decomposition onto $A^{a,b}(X)$. With the convention in the statement, a local $J_t$-holomorphic covector has fixed $J_0$-type expansion $\zeta-t\,\mu\lrcorner\zeta+O(|t|^2)$. Wedge-multiplying over the holomorphic covector slots of the $(p,q)$ component, the terms contributing to fixed type $(p-1,q+1)$ are exactly those in which one holomorphic covector is replaced by its first-order antiholomorphic correction. Each such replacement carries the coefficient from the displayed expansion, namely the minus sign. Thus
\begin{align*}
\operatorname{pr}^{p-1,q+1}_{J_0}(\partial_t\eta_t|_{t=0})=-\mu\lrcorner\alpha
\end{align*}
in Dolbeault cohomology.
The remaining possible terms do not affect this cohomology class. Higher Hodge components of $\eta_0$ have holomorphic degree at least $p+1$, so one first-order replacement can lower them only to holomorphic degree at least $p$, which vanishes in the quotient $F^{p-1}_0/F^p_0$. Changing the local lift or the closed representative changes the de Rham derivative by a cohomologically zero term; after projecting to fixed type $(p-1,q+1)$ this gives a Dolbeault-exact form. Therefore the only surviving class in the quotient is $-[\mu\lrcorner\alpha]$.
Now apply the definition of the second fundamental form. It is
\begin{align*}
\theta_p(\partial_t)([\alpha])=[\nabla_{\partial_t}\sigma(0)]\mod F^p_0.
\end{align*}
In the de Rham realization of the Gauss-Manin connection, $\nabla_{\partial_t}\sigma(0)$ is represented by $\partial_t\eta_t|_{t=0}$. Passing to the quotient $F^{p-1}_0/F^p_0$ keeps exactly the fixed $J_0$-type $(p-1,q+1)$ component. Therefore
\begin{align*}
\theta_p(\partial_t)([\alpha])=-[\mu\lrcorner\alpha]\in H^{q+1}(X,\Omega_X^{p-1}).
\end{align*}
This proves the contraction formula at the level of the second fundamental form.[/guided]