Kura 0.8.4 makes the cache more resilient when replicas restart or churn.
- Freshly restarted replicas are no longer selected as the public primary until they have bootstrapped their artifact set from peers, which fixes cache misses right after a pod restart. Single-replica instances still serve traffic immediately.
- Temporary staging for uploads, replication bootstrap, REAPI ByteStream writes, and multipart assembly now lives on the per-pod persistent data volume under a bounded budget. When the budget is exceeded, Kura returns 503 backpressure instead of failing or evicting the pod.
- The Helm chart exposes the staging budget as
config.tmpDirMaxBytes, defaulting to 8 GiB.