Run an inference cohort

1. Call endpoint-scoped capabilities. available=true means the current account, API-key scope, endpoint assignment, live residency and compatibility registry all admit the endpoint at that moment.
2. Optionally call cohorts:plan. Planning is read-only and does not reserve memory, move tensors or create a billable root.
3. Call cohorts:run with execution.mode=cohort and an idempotency key.
4. Read the summary and events, or connect to the events URL with Accept: text/event-stream.
5. Cancel with cohort:cancel if the result is no longer needed.

Never reuse one idempotency key for different input or policy. A retry with the same request is safe; a conflicting fingerprint returns an explicit conflict.

See the REST, Python, and MCP examples.