Use an explicitly selected LiteLLM profile, configured admin/spend endpoint,
and secret-store credential. Validate the deployment-specific response and map
only bounded spend/budget information into the provider-usage contract.
Missing admin access remains unavailable rather than being guessed from normal
OpenAI-compatible inference responses.
Verification
cd apps/backend && go test -tags fts5 ./internal/agent/usage ./internal/backendapp