Clarification on use of attributes vs datasets for descriptions in neurodata types

[ehennestad]

I’ve noticed that descriptions of neurodata types are sometimes specified using attributes (e.g., a description attribute) and other times as datasets. Sometimes, there is an additional quantity key for some description datasets, but not always.

Would it make sense to clarify the intended convention for when a description should be stored as an attribute vs a dataset?

For example:

Is there a size or content complexity threshold?
Is this governed by a general schema design rule?
Are there legacy reasons for the differences?

This clarification could help ensure consistency in extensions and in interpreting the core schema.