The intent in making the coefficients of $r(x)$ small is not to make life hard for the attacker, but to make life possible for the intended recipient.
Recall that $h(X)$ is chosen to be of the form
$$h(X)\equiv \frac{g(X)p}{f(X)}\pmod q$$
where the coefficients of $f(X)$ and $g(X)$ are small and secret, $p$ is a prime small relative to $q$ and $q$ is large relative to all of the small quantities.
Now consider the decryption which starts with multiplication of $c(X)$ by $f(X)$ to give
$$c(X)f(X)\equiv r(X)g(X)p+m(X)f(X)\pmod q$$
if the coefficients of $f$, $g$, $m$ and $r$ are all small then with high probability we can write
$$c(X)f(X)= r(X)g(X)p+m(X)f(X)$$
over the integers without reduction modulo $q$ by rounding coeffcients towards 0 (i.e. coercing into the interval $[-q/2,q/2]$). On the other hand, if $r(X)$ has any large (relative to $q$) coefficients, we almost certainly cannot do this.
If our rounding process is successful then we can remove the contribution of $r(X)$ by reducing modulo $p$ and then recover $m(X)$.
Note that the rounding process is not guaranteed to be successful as the degree grows and so NTRUEncrypt carries a probability of failure dependent on the coefficient size and degree and which we must try to keep low. Note also that the failure can be dependent on the choice of $m(X)$ which can allow an attacker to actively gain information about the private key (again with a probability that we try to keep small).
You are correct that $r(X)\pmod q$ can be recovered by solving a CVP, irrespective of the size of its coefficients (assuming that the coefficients of $m(X)$ are small).
