The reduction is not done to the diagonalization language, but from the diagonalization language. The difference is very important here.
The idea of reduction is that the fact that there is a reduction from $A$ to $B$, noted $A\leqslant B$, can be translated as:
if we know how to solve $B$, then we know how to solve $A$.
Another interpretation could be:
if we cannot solve $A$, then we cannot solve $B$.
The principle of such a reduction is the following: from an instance $x$ of the problem $A$, construct an instance $y$ of the problem $B$ such that $x\in A\Leftrightarrow y \in B$.
In the reduction in HMU, the proof gives a reduction from $L_d$ to $\overline{L_u}$. Since the book has previously proved that $L_d$ is not recursive (indeed, it is not even recursively enumerable), the reduction $L_d\leqslant \overline{L_u}$ proves that $\overline{L_u}$ is not recursive, hence $L_u$ is not recursive.