[guided]The purpose of introducing the path $h_t$ is to turn metric independence into a differential statement in the parameter $t$. If we can show that the $t$-derivative of $c_k(E,h_t)$ is an exact form, then the de Rham cohomology class cannot change as $t$ varies.
For each $t \in [0,1]$, let
\begin{align*}
\nabla_t: \Omega^0(X;E) \to \Omega^1(X;E)
\end{align*}
be the Chern connection of $(E,h_t)$, and let
\begin{align*}
\Theta_t \in \Omega^2(X;\operatorname{End}(E))
\end{align*}
be its curvature. The derivative of the connection in the parameter $t$ is an endomorphism-valued one-form
\begin{align*}
A_t \in \Omega^1(X;\operatorname{End}(E)),
\end{align*}
defined by
\begin{align*}
A_t = \frac{\partial \nabla_t}{\partial t}.
\end{align*}
This is a tensorial object because the difference of two connections on the same vector bundle is an $\operatorname{End}(E)$-valued one-form, and the derivative of a smooth path of connections is obtained as the limit of such differences.
We apply the Chern-Weil transgression formula to the smooth path $(\nabla_t)_{t \in [0,1]}$ and the invariant polynomial $P_k$. The formula requires two hypotheses: first, that $(\nabla_t)$ is a smooth path of connections on a fixed complex vector bundle, and second, that $P_k$ is invariant under conjugation. The first holds because $h_t$ is a smooth path of Hermitian metrics and the Chern connection depends smoothly on the metric. The second was proved from determinant invariance. Therefore
\begin{align*}
\frac{\partial}{\partial t}\widetilde P_k(\Theta_t,\dots,\Theta_t)
=
d\left(k\,\widetilde P_k(A_t,\Theta_t,\dots,\Theta_t)\right).
\end{align*}
Since
\begin{align*}
c_k(E,h_t)=\widetilde P_k(\Theta_t,\dots,\Theta_t),
\end{align*}
we obtain
\begin{align*}
\frac{\partial}{\partial t}c_k(E,h_t)
=
d\,T_{k,t},
\end{align*}
where
\begin{align*}
T_{k,t}:=k\,\widetilde P_k(A_t,\Theta_t,\dots,\Theta_t)
\end{align*}
is a complex-valued differential form of degree $2k-1$. This is the exactness statement needed for metric independence. For $k=0$, no transgression is needed, since $c_0(E,h_t)=1$ for all $t$.[/guided]