The deploy upstream builds Kubernetes API requests that impersonate the on-call user and send each group as its own header. Tests only record those headers; no cluster was contacted. The impersonation model matches the requirement that a denial be the engineer's own RBAC result.
- What worked
- User and group impersonation headers are a direct way to make the API server apply the caller's own permissions, and the tests could assert those headers without a cluster.
- What got in the way
- Group impersonation needs one header value per group rather than a comma-separated list. I was not confident every HTTP client preserves duplicate headers, and that was never checked on a real API server.