[guided]Let $H := \langle S,T\rangle$. The reduction is the Euclidean algorithm written in matrix language. The matrix $T^q$ performs the elementary row operation
\begin{align*}
(a,c)^\top \mapsto (a+qc,c)^\top,
\end{align*}
while $S$ swaps the two entries up to a sign:
\begin{align*}
(a,c)^\top \mapsto (-c,a)^\top.
\end{align*}
Both operations preserve the greatest common divisor because they are invertible integer linear transformations.
We prove the precise reduction statement. Let $(a,c) \in \mathbb{Z}^2$ satisfy $\gcd(a,c)=1$. If $c=0$, then $\gcd(a,0)=|a|=1$, so $a=\varepsilon$ for some $\varepsilon \in \{1,-1\}$, and the identity matrix already sends $(a,c)^\top$ to $(\varepsilon,0)^\top$.
Now suppose $c \neq 0$. Choose integers $q$ and $r$ such that
\begin{align*}
r = a+qc,
\qquad
|r| < |c|.
\end{align*}
This is exactly the Euclidean step: we replace $a$ by a remainder modulo $c$ whose absolute value is smaller than $|c|$. Applying $T^q$ first and then $S$ gives
\begin{align*}
S T^q
\begin{pmatrix}
a \\
c
\end{pmatrix}
=
S
\begin{pmatrix}
a+qc \\
c
\end{pmatrix}
=
S
\begin{pmatrix}
r \\
c
\end{pmatrix}
=
\begin{pmatrix}
-c \\
r
\end{pmatrix}.
\end{align*}
The important feature is that the second coordinate has changed from $c$ to $r$, and $|r|<|c|$. Thus repeated application of this operation must terminate, because there is no infinite strictly decreasing sequence of non-negative integers.
At the terminal stage the second coordinate is $0$. Since every step was multiplication by a matrix in $H$, the total product is some $w \in H$. Since every step is invertible over $\mathbb{Z}$, the greatest common divisor remains $1$, so the terminal vector must be $(1,0)^\top$ or $(-1,0)^\top$. Hence there exist $w \in H$ and $\varepsilon \in \{1,-1\}$ such that
\begin{align*}
w
\begin{pmatrix}
a \\
c
\end{pmatrix}
=
\begin{pmatrix}
\varepsilon \\
0
\end{pmatrix}.
\end{align*}[/guided]