I think the smaller range should be on the order of 1...2^{128}, assuming 128 bit soundness is desired. In the case of 2 signatures, with r_1 = 1, its easy to show that 1...2^{64} only has 64 bits of soundness, and this extends to larger aggregations if an attacker can restrict the signatures to be in a subgroup of order 2^{64}, which could likely exist if the pairing friendly curve used is also a SNARK friendly curve.
Alternatively, it may suffice that a single client may falsely accept with a lower soundness, as long as its unlikely that everyone is fooled, but that seems a bit precarious imo.