Topics discussed:

• Gino Canessa laid out the closed door in R4: there is no conformant way to define your own resource type. You can't put it in a capability statement, you can't write search parameters for it, you can't write an operation definition against it. Every one of those had to be modified for R6, which is precisely why the R6 work exists.
• Wearing the implementer's hat, Gino said the view on the ground is "I don't care if it's conformant, I want an operation on ViewDefinition in R4, because that's what customers are doing" — and at the implementation level nobody does care. The problem is the definitional side: he doubted any committee would agree to publish an HL7-supported IG full of non-conformant artefacts. He wondered aloud about an "R4C" that adds additional-resource support, then admitted he hasn't raised it with Graham "because he might figure out a way to reach through the computer and strangle me".
• John Grimes thought it mattered less than it looks. ViewDefinition isn't a generic resource you add to a server as another data model and you're done — if you implement it at all you're adding special functionality anyway, so first-class status buys less here than elsewhere. His advice: leave it, adopt the mechanism when FHIR has one, move the IG to R6 when R6 lands.
• The concrete blocker was that IG Publisher rejects an operation definition bound to ViewDefinition. Gino's workaround is a system-level operation taking a canonical URL that points at the ViewDefinition, so no resource type is needed. It's the same game played in the Subscriptions backport IG, where there are no SubscriptionTopic resources, only URLs — "if you just happen to say that URL points to a subscription topic, then that's what it means."
• John said Parquet on FHIR needs a second implementation running alongside his before it graduates from interesting to standards-worthy. He also floated rebuilding MIMIC-IV's derived views as ViewDefinitions: mimic-derived is a set of value-added tables researchers reach for because they're more consumable than the raw ones, so publishing them as views over MIMIC-IV on FHIR would exercise the spec and make the whole chain reproducible with standard tooling.