No disagreement on the lighter touch of the benchmark in this case, it was primarily for illustrative purposes. If you have ~ 10 GB of data then all the standard serial index creation should perform just fine. It's the same case more broadly for Citus, if you only have 10 GB of data there's no point in thinking about scaling it out, it's not until you get to 100 GB or larger that you want to consider something beyond a single node or likely need more parallelism.
But if you have a Ferrari, you don't use it to take your child to school in 14 seconds, you drive it on the open road.
Show a case where serial index builds are impractical, but parallel builds bring it back to a reasonable level.
EDIT: this sounded more negative than I intended. Obviously parallel index build is an important feature.