Let $G$ be a finite group and let $V$ be a finite-dimensional vector space over the field $K$. Let $\rho : G \rightarrow GL(V)$ be a representation of $G$. Let $W\subset V$ be a proper nontrivial subspace of $V$ which is stable under the action of $G$, i.e.
$$ \rho(g)w \in W$$
for all $w \in W$ and for all $g \in G$. Then the representation is said to be reducible. If this is not the case, the representation is said to be irreducible. On the other hand, we say a representation is completely reducible if the vector space on which the group $G$ is represented can be written as a direct sum of $G$-invariant subspaces, i.e. $V= \bigoplus_i W_i$, with each $W_i$ being $G$-invariant as defined above.
EDIT: I assume here that the sum has more than one element, so that an irreducible representation is never completely reducible at the same time.
My question: does reducible imply completely reducible? Where can I find a proof? I remember a theorem saying that for any finite group, if a representation is reducible then it is completely reducible, but I can't find it right now. In terms of matrix representations, the question is whether we can ever have
$$\left(\begin{array}{@{}c|c@{}} \rho^{(1)}(g) & b(g) \\ \hline 0 & \rho^{(2)} (g) \end{array}\right) \, ,$$
with $b(g) \neq 0$. If so, why?
EDIT #2: I think Maschke's theorem is what I am looking for: if the characteristic of the field is zero, then what I said above is true.