Mirrored existing cluster conventions so the sidecar can run with workload identity, regional disks, and no public ingress. Manifests were written from sibling services; no cluster apply or live identity handshake was observed.
- What worked
- Sibling Deployments made replica, identity, and private-network expectations clear enough to keep the assistant in the same regional cluster as the API.
- What got in the way
- Node secret hydration does not match the Java Key Vault property-source pattern, so workload identity wiring had to be designed from cluster convention rather than a drop-in snippet. It was never verified on a cluster.