fix: eliminate 502 errors during rolling deployments

[scotwells]

Summary

Clients occasionally see brief 502 errors when the API server is updated. This PR makes deployments seamless so clients never see errors during rollouts.

What changed:

• The gateway now automatically retries failed requests on a healthy pod instead of returning the error to the client
• Health checks detect pods shutting down within seconds and stop sending them traffic
• The API server signals it's going away before actually stopping, giving the gateway time to react
• A disruption budget prevents too many pods from going down at once during cluster maintenance
• The rollout strategy now waits for a new pod to be fully ready before terminating the old one

Closes #559

Test plan

• Deploy to staging with 2+ API server replicas
• Trigger a rolling update (image tag change) while sending continuous API requests
• Verify zero 502 errors are returned to clients during the rollout
• Verify the BackendTrafficPolicy is accepted by Envoy Gateway (kubectl get btp -n milo-system)
• Test a voluntary node drain to confirm the PDB keeps at least one pod available

🤖 Generated with Claude Code

[joggrbot]

📝 Documentation Analysis

All docs are up to date! 🎉

---

✅ Latest commit analyzed: 46f25f2 | Powered by Joggr

[ecv]

hahaha