[Jonathan-Eid]

Bumps the worker spread maxSkew from 2 to 3 so Karpenter packs 3 workers per node instead of 2.

• With a hard hostname spread, Karpenter plans each new node with maxSkew workers. At 2, 1024 workers need 512 nodes.
• Prod only got 3 per node (384 nodes) because the catchup nodepool's 3072 cpu limit capped it. stellar/kube#5513 added fallback pools with their own limits, so build 1963 went to 512 nodes: $239/hr and $732 for catchup vs ~$635 on 384 nodes.
• m8a.2xlarge fits 3 workers by memory (3 × 8Gi of ~28.7Gi allocatable), so 3 fills the node.

Testing (ssc-test, prod's catchup pools mirrored, 60 pods with the worker's requests and spread rule):

• maxSkew: 2: 30 nodes, 2 per node
• maxSkew: 3: 20 nodes, 3 per node, 84% memory requested
• helm template renders maxSkew: 3 on every worker pod

🤖 Generated with Claude Code