Kura runner-cache failures now fall back to the public cache instead of timing out

When a Kura cache node’s readiness probe started returning 503 (for example, during a wedged cross-region replication mesh), builds kept being routed at that dead node-port endpoint and timed out on module-cache downloads instead of degrading. Dispatch now checks a readiness heartbeat (stamped by the reconciler, kept fresh for about 120 seconds) before handing builds the private endpoint; once the heartbeat goes stale, the build falls back to the public cache. CI and customer macOS runner builds now degrade gracefully during a Kura outage rather than timing out.