The basic question is how any authority serves to identify a particular entity. A name string alone is not enough. A declared relationship to a title is better, especially if the title includes subject words and a publication date relevant to the entity. More generalized information of the type now included in RDA authorities provides additional support for selection and identification. Best (though never perfect) is when a skilled person has reviewed an authority and confirmed that it satisfactorily identifies a specific entity.
The BIBFRAME authority has a name string and a link to a bibliographic description, so it implies a relation to a title; and the bib description may include a role as part of its link to the BF authority, another bit of identifying information. By itself, the BF authority would be inadequate for identification purposes, but by extension, including information from the bib descriptions it links to, it could suffice. When a BF authority also links out to an external authority (LCNAF, ISNI, whatever), any information that authority can provide becomes part of this extend set of identifying information.
All this depends on treating the BIBFRAME authority as more than just a name string. If the bib descriptions to which a BF authority is linked internally are related to different entities rather than just one entity, then the BF authority has no value for identification purposes. It's hard to see how any automated processing of a large database could reliably generate useful BF authorities based on unqualified and unauthorized name strings from bib descriptions (of which our large database has many). If the same name string appears on more than one bib description, either too often different entities with undifferentiated name strings will be put on a single BF authority, or too often multiple BF authorities for the same entity will be created based on the name's appearance in multiple bib descriptions.
If automated conversion of MARC records is part of the BIBFRAME plan, and if automated generation of BF authorities is included in that conversion, then we may want a way to indicate whether a BF authority is regarded as identifying a specific entity uniquely, or as identifying a specific entity but maybe not uniquely, or only as authorizing a common name and no identity. Either that, or we skip BF authorities in some cases and include unauthorized name strings in BF bib descriptions.
One of the challenges going forward is formulating and operationalizing brief presentations of this extended authority data which can do a better job of supporting the identification function than our unique name strings have done; but that's probably more for implementing systems to develop than for BIBFRAME itself.
Responding Kelley's original questions: URI's have value as authorities when they can return sufficient information to identify an entity. That wouldn't need to include a unique name string, but for many purposes would need to include some kind of preferred name and be human-readable.
I like Kelley's suggestion that formal role identification be kept at a more general level with the more unusual role labels retained as part of the bib description, but not as new relators. This reminds me a bit of the MARC language codes, where more terms than codes are listed since some languages are assigned only language group codes; except that the number of unique relator terms is, on Kelley's evidence, infinitely expanding.