My guess? They're giving an estimation based on talking to a few shards of a much larger sharded system rather than trying to actually get canonical results for

My guess? They're giving an estimation based on talking to a few shards of a much larger sharded system rather than trying to actually get canonical results for every shard - since it's unlikely that you need a precise count across that many repositories (which would be really expensive to calculate in real time).