In his paper Constructive Renormalization Theory, V. Rivasseau describes the idea of Wilson's approach of solving path integrals step by step. In section 1.4, page 5, however, there is a statement which I do not follow. He discusses a $\phi^{4}$ theory and decomposes the covariance into a sum: \begin{eqnarray} C_{0}(p) = \int_{1}^{\infty}e^{-\alpha(p^{2}+m^{2})}d\alpha \quad \mbox{and} \quad C^{j}(p) = \int_{M^{-2j}}^{M^{-2(j-1)}}e^{-\alpha(p^{2}+m^{2})}d\alpha \tag{1}\label{1} \end{eqnarray} in such a way that: \begin{eqnarray} C_{\rho}(p) = \sum_{j=0}^{\rho}C^{j}(p) \tag{2}\label{2} \end{eqnarray} He proceeds to explain how the partition function can be obtained by performing interativelly $\rho+1$ (convolution) integral. At some point, he states the following:

We see that constructing the ultraviolet limit is the same as finding a bare action $S_{\rho}(\phi)$ such that the effective action, or renormalized action $S_{0}(\phi)$ converges as $\rho \to \infty$.

I might be missing something, but I don't fully understand this statement. First, it seems that, in the limit $\rho \to \infty$ one recovers the the regularized theory, because (\ref{1}) is a telescoping series with UV cutoff and as $j \to +\infty$ one ends up with regularized covariance \begin{eqnarray} C_{0}(p) = \frac{1}{p^{2}+m^{2}}e^{-(p^{2}+m^{2})}. \tag{3}\label{3} \end{eqnarray} In other words, I don't see how it is possible to take the UV limit using this decomposition into steps as a limiting case. Second, it is a little bit odd for me to say that one finds $S_{\rho}(\phi)$ so that $S_{0}(\rho)$ exists when $\rho \to \infty$. The flow is going backwards, from $\rho$ to zero, how can you then $\rho \to \infty$ afterwards? Is it just because $S_{0}$ should depend on $\rho$ somehow?

Add: A related question is the following. I think the main idea is to proceed with the integrating steps until one achieves a fixed point. What does one expect of this fixed point? I mean, is it dependent on the UV cutoff so one has to deal with removing this cutoff or does one expect that it is independent on this cutoff so the removal is trivial?

In short, I am really confused about how one plans to remove the cutoff. I am not convinced that this can be done by a limiting case of the RG iteraction as Rivasseau stated. However, if this limiting case is achieved and it does depend on the cutoff, how to remove it? I know these topics are really complicate in practice and there is no recipe, but I would be happy with a picture of what one is trying to do or expect to have. In other words, what is the philosophy?