Is there any reason (counterexample) why the composition of a monad with itself should not be a monad again? It appears to me that the distributive law is in this case just the identity (on the composition).
Given two monads $T$ and $S$ over the same category we can compose them whenever we have a distributive law at our disposal, that is, a natural transformation $d:TS\rightarrow ST$ satisfying certain commutative diagrams.
With it, we have a new monad structure for $ST$ whose unit is the composition of units $\eta_S \eta_T$ and whose multiplication is $\mu_S \mu_T SdT$.
My guess, and that's my question, is that for $TT$ we also have a new monad structure whose unit is $\eta_T \eta_T$ and whose multiplication is $\mu_T \mu_T$ and where d is the identity on $TT$.
Is that right? If not, please give a counterexample.
Just check the monad laws. Using whiskering instead of horizontal composition, so $F\tau = id_F*\tau$ and $\tau_F = \tau*id_F$. Call $\eta_T$ just $\eta$ because otherwise there's a notational conflict, and similarly for $\mu_T$. So $\eta : Id \to T$ and $\eta_T : T \to T^2$. The monad laws are $\mu \circ \eta_T = id_T= \mu \circ T\eta$, and $\mu \circ \mu_T = \mu \circ T\mu$.
$\eta * \eta = \eta_T\circ \eta = T\eta \circ \eta$. Similarly, $\mu*\mu = \mu_T \circ T^2\mu = T\mu \circ \mu_{T^2}$. Further, $(\eta*\eta)_{T^2} = \eta_{T^3}\circ\eta_{T^2} = T\eta_{T^2}\circ\eta_{T^2}$ and similarly $T^2(\eta*\eta)=T^2\eta_T\circ T^2\eta = T^3\eta\circ T^2\eta$. Likewise for $\mu$. Note, the counter-example from the Composing Monads paper mentioned is for $\mu_T\circ\mu_{T^2}$ which is not $\mu*\mu$.
We need to prove $(\mu*\mu)\circ(\eta*\eta)_{T^2} = id_{T^2} = (\mu*\mu)\circ T^2(\eta*\eta)$ and $(\mu*\mu)\circ(\mu*\mu)_{T^2} = (\mu*\mu)\circ T^2(\mu*\mu)$. Now we calculate: $$\begin{align} (\mu*\mu)\circ(\eta*\eta)_{T^2} & = T\mu \circ \mu_{T^2}\circ T\eta_{T^2}\circ\eta_{T^2} \\ & = T\mu \circ \eta_{T^2} \end{align}$$ And we're stuck. If you draw out the 2-cells as a diagram it's easy to see what goes wrong. Both $\eta$s interact with only one of the $\mu$s which can only cancel one of them. This leaves an $\eta$ and a $\mu$ that don't interact, and thus can't cancel out to the identity. Alternatively, if you use string diagrams the issue is also self-evident.
To use the free monoid monad (the List monad) as a counter-example because it's convenient, we see that $(\mu*\mu)\circ(\eta*\eta)_{T^2}$ produces a singleton list of a flattened list. This is even easier to see by using naturality of $\eta$ in the last equation above to get $T\mu\circ\eta_{T^2} = \eta_T\circ\mu$. For example, $((a)(b))\mapsto ((ab))$ which is clearly not the identity. A similar thing will happen for just about any algebraic structure. Below is a formal proof in Agda that the List monad is a monad and that horizontally composing the unit and multiplication don't give a monad.