Fleet-level re-pinning on sustained cluster degradation #105

What problem are you facing?

Once a ModelReplica is pinned to an InferenceCluster, modelplane intentionally retains the pin across cluster degradations. A sensible default for transient outages.

But sustained degradations have no escape valve. When a node is out for days, a pool is permanently shrunk, a GPU class is deprecated, or hardware never comes back, the placed replica stays stuck on the degraded cluster. The ModelDeployment reports ready < total indefinitely. The only recovery available to the operator is to delete the ModelDeployment and recreate it .

How could Modelplane help solve your problem?

When a placed ModelReplica's cluster has an effective-capacity shortfall that persists past a threshold, modelplane should re-pin the replica to another eligible cluster. Transient blips continue not to trigger churn.

The "how" is open — including what counts as "sustained," whether re-pinning is opt-in or default, and how draining/eviction semantics interact with serving traffic. What matters is that re-pinning becomes a decision the platform can make, instead of something only the operator can force by deleting and recreating the deployment.

Related: #106 (cross-provider normalization of the capacity signal that drives this decision).