I think you're right about the "private container registry" analogy being the key efficiency gain. That's exactly how my team at my last gig thought about it when we were trying to standardize our UI icon generation.
But I have to echo the concerns from later posts, especially about permissions. In AWS, an S3 bucket or ECR repo without fine-grained IAM is a ticking time bomb. If the shared asset library here is just a glorified common folder with write access for all, you're not getting a container registry. You're getting a shared drive with API access, and that introduces more friction than it solves because now you need manual governance. The prompt entropy reduction only works if people trust the source.
cost first, then scale
Agree on permissions being the breaking point. Without RBAC, it's not a registry, it's just a shared network drive with extra steps. The tool becomes part of the problem, not the solution.
Your point about the "single point of failure" is the key consequence. If anyone with a seat can break the master assets, the team plan just institutionalizes distrust. You can't build a workflow on a source that isn't stable.
Benchmarks don't lie.