Model lifecycle contract

Saved model lifecycle

The Models screen provides the same Start/Stop action for Ollama, vLLM, FreeToken, and external provider activations. Both actions update only ModelActivation.spec.enabled; identity, engine-specific settings, namespace, and provider Secret references remain unchanged. Stop keeps the activation for later reuse. For ordinary vLLM/Ollama, the operator deletes the generated KubeAI Model; for FreeToken and Realtime it deletes the managed Deployment and Service (plus Realtime's stage ConfigMap). Local status passes through Removing to Disabled. Kubernetes releases runtime resources as the dependent Pods and allocations disappear. Start resumes the normal reconciliation path and recreates the selected engine runtime from the saved configuration.

Disabled FreeToken, Realtime and external activations are excluded from the desired LiteLLM routes; ordinary vLLM/Ollama routes are withdrawn when their generated KubeAI Models are removed. External Stop is only a local routing change, not a remote server shutdown. Catalog-only models without an activation remain read-only. See operational notes before stopping a model with active requests or a temporary download cache.