I thought a little about this question and I'm still not a 100% sure how to answer it. However, I think the crucial point about this assumption is that (with the assumption) the angular derivative of $\phi$ at "infinity" is always proportional to the action of a Lie algebra element of $\mathfrak{g}$, i.e. $$\partial_\theta \phi= i \lambda^a T_a \phi$$ is satisfied asympotically at $r\rightarrow\infty$ with $\alpha^a$ depends on $\theta$. This is because $G$ has a natural transitive action on the quotient space $G/H$, i.e. given points $p,q\in G/H$ there is always a $g\in G$ such that $g\cdot p=q$. Thus, we can choose a fixed point in $G/H$, call it $\phi_0$, and write the asymptotic behaviour of $\phi$ as $$\phi(\infty,\theta)= g(\theta)\cdot\phi_0 = \exp\left(i\alpha^a(\theta) T_a\right)\cdot \phi_0.$$ Only if this is the case, the partial derivative can be compensated as a gauge of $A$ and, thus, the covariant derivative vanishes fast enought.
Now, what happens if we do not assume the space of minima to be $G/H$? Let's denote the space of minima by $Z$ and as above take a $\phi_0\in Z$. We will seperate this in two cases:
The connected component of $\phi_0$ in $Z$ is $G/H$: In this case nothing changes and the argument is still valid. Note, that given a minimum $\phi_0\in Z$ always gives rise to a whole $G/H$ subspace in $Z$ since $U$ is supposed to be $G$-invariant and thus $G\cdot\phi_0 \cong G/H \subseteq Z$.
The other case: If the connected component, call it $Z_{\phi_0}$ of $\phi_0$ is not $G/H$ we have to be more careful. In this case, the field $\phi$ has additional asymptotic degrees of freedom. It can oscillate between different $G/H$ subspaces. To illustrate this, imagine a symmetry breaking pattern $SO(3)\rightarrow SO(2)$, then the quotient is the 2-sphere $SO(3)/SO(2)=S_2$. The space $Z$ could for example be equal to $I\times S_2$, with $I$ some interval. For a finite action/energy we assume $\phi(x)\in Z$ assymptotically. The movement of $\phi$ in the $S_2$ subspaces generated by the $G$-action can be compensated by the gauge field. However, this is not the case if $\phi$ moves in $I$ as $\theta$ varies. Thus, any change of $\phi$ in the $I$-component results in an infinte contribution to the kinetic energy.
Anyway, all the standard symmetry breaking potentials I have seen so far had the property that $Z=G/H$. We could in principle allow for potentials as in case 1 above. But I don't think it will be useful for anything.
As a conclution, I think the assumption is done to ensure that the change of $\phi$ "at infinity" can be compensated in the gauge field $A$. Without the assumption, this is not true in general. However, I think there are still solutions with finite energy, but we propably get trouble when trying to quantize the theory. But this is just speculation from my side.
Although it is not a full answer to your question, I hope these thoughts might help you!