Thanks for the quick reply.
By looking at examples, it seems to me that the correct genotype for each sample could be computed after joining. However, this may require computing new set of alleles in the ALT field (not just simply put them all) and all genotypes must be recomputed too. I do agree that finding correct genotypes after joining these sites require a complex logic to be implemented. But considering that genotype is the most important information and is used in many analysis, it may worth to be considered.
I am not well familiar with the statistical information in VCF files (PL, AD). There might be the case that we cannot recompute them after joining. In this case, they can be removed from the dataset and another flag could indicate removal.