The metadata is not related to the read groups.

As the skywarrior person said in the post you linked:

BWA hardclips reads if there is a significant discordance between the best matching kmer and the read. These hardclips may end up costing you a particular structural variant or a true indel call. Merging unmapped bam and initial alignment restores the hardclips which I know of no solution for that in BWA parameters.

Thus you are not really losing metadata... you are potentially losing actual data from your original sequencing reads. This step may be unnecessary depending on the type of dataset you have (Exome vs. Whole genome) or furthermore maybe you don't care about certain structural variants and/or know that they aren't present in your dataset.