The generic model for a MAC is the following: the attacker is given access to a block box which implements the $S$ function with a key $k$ that the attacker does not know of. The attacker is allowed to make $q$ requests to the box on messages that he can choose arbitrarily. The goal of the attacker is to make a forgery, i.e. produce values $m$ and $t$ such that $V(k,m,t) = \mathbf{true}$ but $m$ is distinct from all the $q$ messages sent as requests. In other words, the attacker obtains $q+1$ valid message+MAC pairs from only $q$ requests.
Now, suppose that a forgery is possible on $S^{big}$. This means that the attacker can send $m_1$, $m_2$,... $m_q$ to the box implementing $S^{big}$, obtain $t_1$, $t_2$,... $t_q$, and somehow compute $m$ and $t$ where $m$ is distinct from all the $m_i$, and $V^{big}(k,m,t) = \mathbf{true}$. Then look at $H(m)$ and compare it to all the $H(m_i)$; exactly one of the two following must hold:
- If $H(m) = H(m_i)$ for some $i$, then, by construction, the attacker found a collision on $H$.
- If $H(m) \ne H(m_i)$ for all $i$, then the attacker implicitly made $q$ requests to $S$ (for all $H(m_i)$) and then computed $m$, which means that $H(m)$ and $t$ are such that $V(k,H(m),t) = \mathbf{true}$. The attacker then succeeded at a forgery attack on $S$: he computed a pair containing a short message ($H(m)$) and a token $t$ which are accepted by $V$, and yet at no point was that exact $H(m)$ sent to the box implementing $S$.
It follows that if $S$ is strong against forgeries and $H$ is collision-resistant, then $S^{big}$ must also be strong against forgeries.