There are indeed some immediate lessons and you’ve highlighted them here. I’ll get into more specifics.
It was new to me that OpenRefine reconciliation endpoints can receive properties and values that may be used to enhance the likelihood score on a term being reconciled – this is a significant feature and it does deserve careful consideration. In particular, it will force us to think very carefully about types and properties (what you called attributes) from the perspective of a user who needs to turn strings to things and in what context they are working. Are they resolving a reference list and want a linked & citable reference to a collection? Are they wanting a linked collection code for a collection? Are they wanting a collection’s parent institution? Are they wanting the collection’s name in an alternate language?
Wikidata, understandably, has a robust reconciliation endpoint and if I’m not mistaken, is available in OpenRefine at install. It would be in our best interest to model what is a collection in wikidata. The immediate advantage of doing so is access to users who can verify that the reconciliation endpoint that results from this work suits their needs.