Renormalization is a coarse graining procedure where every step involves "forgetting" some high frequency modes. If the partition function you refer to is the original one that includes the forgotten modes, then you can say the thermodynamic properties don't change. However, the properties to which you are sensitive definitely do change (unless you are already at a fixed point). So the interesting partition function to consider is an effective partition function for the new scale. Wilson and friends taught us that this effective description looks like the original one with an important difference: the couplings have been redefined. Perhaps this is what you were trying to say :).
The phenomenon you mention of all points in coupling space approaching a fixed point is a good rule of thumb but there are exceptions because RG flows can have limit cycles. Assuming we are not in this exotic situation, which fixed point we reach can be determined by considering the unstable directions which are associated with the so called relevant operators. The typical example, present in most statistical systems, is the reduced temperature. If $T > T_c$, RG steps make the effective temperature rise to a larger value. Experimentally, this corresponds to "losing sight" of domains and their typical size as you zoom out leading to a $\xi = 0$ phase as you mention. However, when $T = T_c$, RG steps will leave you there. The diminishing effect of irrelevant operators means that the system's behaviour will stabilize onto something non-trivially scale invariant as you zoom out. This has $\xi = \infty$ meaning there are clusters of all sizes and this is what we call a critical point.
A classic book about this stuff is Scaling and Renormalization in Statistical Physics by Cardy. But there are also plenty of issues which are too new to be in books such as anomalies which can sometimes prevent a disordered phase from existing.
