With 900m/1000m CPU limits used, rolling updates fail because the
new pod cannot be created alongside the old one. Setting maxSurge=0
and maxUnavailable=1 kills the old pod first, avoiding quota exhaustion
at the cost of brief downtime during deploys.