If two SNPs are co-inherited 100% of the times, D' prime will always be 1, while r2 incorporates the information about allele frequencies and thus will be lower than 1 if allele frequencies deviate from 0.5 (for example, in case of a perfect disequilibrium, if major allele frequency is 50%, and minor allele frequency 1%, r2 will drop to 0.01). Doesn't this mean that r2 is useful only when searching either for SNPs in LD or for a proxy SNP, in order to prevent us from selecting a SNP that is not common in the population (possibly a rare variant)?
It is not more or less correct except in relation to a given question or methodology.
For instance, there are a lot of methods that make use of r^2 to give information about other things of interest. The PAINTOR algorithm, for instance, uses an LD matrix consisting of r values, rather than D' values, to prioritize candidate causal variants.
This is a case when we might use r or r^2 but it has nothing to do with the rarity of the variant.
Thus, the question as phrased is too broad. It could be correct or incorrect to use either D' or r^2 in certain situations, and those may or may not have to do with rarity of variation.