The addition of angular momenta is associative in the sense that the operators themselves, as operators, are the same, i.e. that the operators
$$
\mathbf J = (\mathbf L_1 + \mathbf L_2) + (\mathbf S_1+\mathbf S_2) = (\mathbf L_1+\mathbf S_1)+(\mathbf L_2+\mathbf S_2)
$$
are the same. However, when we refer to the addition of angular momenta in quantum mechanics, we mean quite a bit more than that; more specifically, we refer to a re-diagonalization process to find the eigenstates of the total angular momentum operator, and this is not an associative operation.
In the end, both the LS and JJ coupling schemes have the same final goal: to get an eigenbasis for $J^2$ and $J_z$ (and also of $L_1^2$, $L_2^2$, $S_1^2$ and $S_2^2$), but they achieve more than that:
- LS coupling produces a joint eigenbasis of $J^2$ and $J_z$ which is also an eigenbasis of $L^2$ and $S^2$.
- JJ coupling produces a joint eigenbasis of $J^2$ and $J_z$ which is also an eigenbasis of $J_1^2$ and $J_2^2$.
The reason this is possible is that $J^2$ and $J_z$ commute with all of $L^2$, $S^2$, $J_1^2$ and $J_2^2$. However, they're not both simultaneously possible, because neither of $L^2$ or $S^2$ commutes with either of $J_1^2$ and $J_2^2$.
To see this explicitly, you first expand out the commutator:
\begin{align}
[L^2,J_1^2]
& =
\left[ (\mathbf L_1 + \mathbf L_2)^2 , (\mathbf L_1 + \mathbf S_1)^2 \right]
\\ & =
\left[ L_1^2 + L_2^2 + 2\mathbf L_1 \cdot\mathbf L_2, L_1^2 + S_1^2 + 2\mathbf L_1 \cdot\mathbf S_1 \right]
\\ & =
4 \left[ \mathbf L_1 \cdot\mathbf L_2, \mathbf L_1 \cdot\mathbf S_1 \right]
\\ & =
4 \sum_{k,n} \left[ L_{1,k} L_{2,k}, L_{1,n} S_{1,n} \right]
\\ & =
4 \sum_{k,n} \left[ L_{1,k}, L_{1,n} \right]L_{2,k} S_{1,n}
,
\end{align}
and here is the rub: each of those squares includes a bunch of terms linear in the components of $\mathbf L_1$ on either side of the commutator, and those stack up into a single commutator of the components of $\mathbf L_1$ with all the other components of $\mathbf L_1$, and this commutator is not zero.
(If you're really curious, the full commutator evaluates to $[L^2,J_1^2] = 4 (\mathbf L_2 \times \mathbf S_1) \cdot \mathbf L_1$.)
So, what's one to do in this situation? If you're adding up four angular momenta as
$$
\mathbf J = \mathbf L_1 + \mathbf L_2 + \mathbf S_1+\mathbf S_2,
$$
what you'd ideally want is a joint eigenbasis of
- the total angular momentum $J^2$ and one of its components, $J_z$,
- all of the (squares of the) individual angular momenta, $L_1^2$, $L_2^2$, $S_1^2$ and $S_2^2$, as well as
- all of the (squares of the) intermediate combinations, like $L^2$, $S^2$, $J_1^2$ and $J_2^2$, as well as the intermediate-coupling operators like $(\mathbf L+\mathbf S_1)^2$ and its ilk.
The first two goals are perfectly achievable but, because the operators in the third bullet point don't all commute with each other, we need to choose a finite subset of such operators to co-diagonalize with $J^2$ and $J_z$. This is the choice that's embedded in the dichotomy between the LS and JJ couplings.