Also agree strongly - from the use cases that we’ve looked at in for the TDWG Collection Descriptions Data Standard task group, the definition of a natural history collection varies widely depending on the context. The most specific we can get that fits all of them is a fairly generic ‘a group of physical collection objects with one or more common characteristics’, with those characteristics being defined by the context.
It may be that only the more traditional concept of a ‘collection’, such as those that represent a whole institution’s collection or tier below (e.g. herbarium, palaeo collection) need a human-readable identifier due to their historical use in previous and current registries. More granular collections like @elyw mentions may not have the same need. However, every collection needs a machine-readable identifier, and that would have to be unique in a global catalogue.
I think the main trap to avoid (as with specimen identifiers) is conflating the purpose and requirements for human and machine-readable identifiers. The human versions need to short, memorable and informative, and ideally would be globally unique and persistent to avoid confusion, but it won’t break things thing if they aren’t, or if they don’t exist for a collection. The machine versions need to be globally unique, persistent and resolvable, and the ones relied upon by software for unambiguous identification and data linkage.The issues tend to arise when we try to rely on the human-readable identifiers for that, or force human readability concerns on machine identifiers.