While looking at our indexmap dependency, I wondered whether a hash table per element is not too heavy weight for what we are doing. For looking up attributes by name, we can also use a sorted list to get O(log n) instead O(<1>) which considering the typical number of attributes on an element should be good enough. Similarly, adding attributes one-by-one after the fact should be rare, so shuffling them around should be rare as well. What this yields is a more compact representation and simpler machine code (even though the source code gets a bit messier).