A clog doesn't just lose cells, it biases which cells you got - larger nuclei and debris block first, so your 1,500 are a compositionally skewed subset, and the extra time on the instrument means more ambient RNA and more stressed nuclei. So it depends on the analysis: if you're comparing cell type composition across donors, merging a biased 1,500 into a clean 5,500 corrupts the very thing you're measuring and I'd just discard the original. If you only need cells for annotation or per-cell-type expression, merging is fine.

If you do merge, do it at the count matrix level with run kept as a batch covariate, not by pooling fragments. Two libraries means two emulsions, two ambient pools and two sequencing batches, and pooling fragments as though it were one library hides all of that.

One trap specific to your setup: Vireo's donor0/donor1 labels are arbitrary per run unless you give it a reference genotype VCF. Donor0 in the repeat isn't necessarily the same person as donor0 in the original. Either anchor both runs to a genotype reference, or match them afterwards using the genotypes Vireo writes out - otherwise you can silently merge one donor's cells into the other's.