Troubleshoot fallback and budgets

Symptom/code	Meaning	Safe action
preview_not_enabled	The authenticated account/endpoint is not explicitly enabled for Cohort.	Use Classic or enable the compatible causal LLM endpoint through the authorized product control.
insufficient_scope	A restrictive key lacks inference:execute.	Issue a least-privilege key with that scope.
endpoint_not_in_rollout	The endpoint has no active consented assignment.	Do not retry as another endpoint.
no_validated_representation_space / no_safe_cohort	Compatibility, bridge generation or live residency is absent.	Use Classic; do not force a hardware route.
cohort_temporarily_unavailable	A server flag, fabric health or control dependency is down.	Retry with backoff or use Classic when policy permits.
cohort_rate_limited	Account admission budget is exhausted.	Respect Retry-After when present.
idempotency_conflict	The key was reused with a different fingerprint.	Create a new key for the new request.
latency_budget_exceeded	The durable deadline ended the Cohort attempt.	Inspect fallback status; do not wait on old members.
Gateway/node reconnect during dispatch	Delivery is unacknowledged; the exact attempt lease is released and the member remains ready.	Restore the private WebSocket route; the safety drainer resumes automatically within the request deadline.

Cost and latency limits are server-bounded. Settlement is reported explicitly; a zero charge in a test or preview account is not a general pricing commitment.