This note is in Hasura.RQL.DDL.Schema.Cache. It is referenced at:
After building the schema cache, we prune InvalidationKeys for objects that no longer exist in the schema to avoid leaking memory for objects that have been dropped. However, note that we don’t want to drop keys for objects that are simply inconsistent!
Why? The object is still in the metadata, so next time we reload it, we’ll reprocess that object. We want to reuse the cache if its definition hasn’t changed, but if we dropped the invalidation key, it will incorrectly be reprocessed (since the invalidation key changed from present to absent).