[guided]The goal of this step is to show that no partition of the concatenated path can collect more length than the length already present in the two pieces. Let
\begin{align*}
P=\{0=t_0<t_1<\cdots<t_m=1\}
\end{align*}
be a finite partition of $[0,1]$. The only point where the formula for $\alpha=\eta*\gamma$ changes is $1/2$, so we first make sure the partition sees that point. If $1/2$ is not a point of $P$, define $P'$ by inserting $1/2$ into $P$; if $1/2\in P$, define $P'=P$.
Why is this refinement allowed? Suppose a point $r$ is inserted between two consecutive partition points $a<r<b$. The triangle inequality gives
\begin{align*}
d(\alpha(b),\alpha(a))\leq d(\alpha(r),\alpha(a))+d(\alpha(b),\alpha(r)).
\end{align*}
All other terms in the variation sum are unchanged, so inserting $r$ cannot decrease the variation sum. Applying this to the insertion of $1/2$ gives
\begin{align*}
V(\alpha,P)\leq V(\alpha,P').
\end{align*}
Now list the points of $P'$ on the first half of the interval as
\begin{align*}
0=a_0<a_1<\cdots<a_r=\frac{1}{2}
\end{align*}
and the points of $P'$ on the second half as
\begin{align*}
\frac{1}{2}=b_0<b_1<\cdots<b_s=1.
\end{align*}
The first half of $\alpha$ is the path $\gamma$ traversed at twice the speed, so we rescale the first-half partition by defining
\begin{align*}
A:=\{2a_0,2a_1,\dots,2a_r\}.
\end{align*}
This is a finite partition of $[0,1]$ because $2a_0=0$, $2a_r=1$, and the order is preserved by multiplication by $2$. Similarly, the second half of $\alpha$ is the path $\eta$ traversed after the affine change of variables $u=2t-1$, so we define
\begin{align*}
B:=\{2b_0-1,2b_1-1,\dots,2b_s-1\}.
\end{align*}
This is also a finite partition of $[0,1]$.
Using the definition of the concatenation on each half, the variation sum along $P'$ splits exactly:
\begin{align*}
V(\alpha,P')=V(\gamma,A)+V(\eta,B).
\end{align*}
The equality includes the joining point because $\gamma(1)=y=\eta(0)$, so the endpoint of the first piece and the starting point of the second piece agree. Since $A$ is a partition of $[0,1]$, its variation sum for $\gamma$ is bounded above by the supremum defining $L(\gamma)$:
\begin{align*}
V(\gamma,A)\leq L(\gamma).
\end{align*}
Likewise,
\begin{align*}
V(\eta,B)\leq L(\eta).
\end{align*}
Combining these inequalities gives
\begin{align*}
V(\alpha,P)\leq V(\alpha,P')=V(\gamma,A)+V(\eta,B)\leq L(\gamma)+L(\eta).
\end{align*}
Because $P$ was arbitrary, taking the supremum over all finite partitions $P$ of $[0,1]$ proves
\begin{align*}
L(\alpha)\leq L(\gamma)+L(\eta).
\end{align*}[/guided]