elaborate on first class groups property alternative.

Author: cliffhall

seps/2084-primitive-groups.md

@@ -205,7 +205,7 @@ This specification proposal was selected for its ease of understanding since it
 
 - **Primitive's group list passed in a first class `groups` property:** A variation of the proposed specification, but the list of groups to which a primitive instance belongs would be presented by a `groups` property added to the top level of each primitive's schema.
   This idea was discarded because it could lead to backward compatibility issues. For instance, if a server returned a tool, resource, etc, with this property to an older client which validated it against a strict schema that did not contain this property, it would most likely cause an error.
-  Since this proposal spans all primitives, such a compatibility failure would be unacceptable. Consequently, we settled on using a reserved metadata key to pass the primitive's group list.
+  Since this proposal spans all primitives, such a compatibility failure would be unacceptable. We could require SDK developers to implement a "feature flag" that suppresses the top-level groups field in all primitives after protocol version negotiation takes place if there is a mismatch. However, that would be more effort and complexity than simply using a reserved metadata key to pass the primitive's group list, which older clients would simply ignore.
 
 - **Primitive's group list as an array of Group instances not names:** A variation of the proposed specification, but the schema would reference the Groups definition instead of declaring a string (group name). This means that full Group instances would appear in the primitive's group list, significantly increasing the token count when passed to an LLM without modification. Also, beacuse groups can be hierarchical, every child of a given group would carry a duplicate of the parent instance. There was discussion of mitigating the duplication on the line using libraries that perform a marshalling on send/receive, replacing the parent with a pointer to a single copy of the parent instance. This would put an unnecessary burden on SDK developers for no clear benefit, when a client can easily look up a group by name in its cached `groups/list` result.
   More information on this approach can be found in @scottslewis's [proposal](https://github.com/modelcontextprotocol/modelcontextprotocol/discussions/1567).