cilium-health appears to be enabled by default in ENI mode, but there are easily reproducible issues in multinode clusters where the cilium-health instances on some nodes are unable to communicate with simulated endpoints on other nodes.

We should investigate these, and:

• Determine whether there is a bug in the cilium-health logic
• If so, fix it; if not, document here why not with proof;
• And/or consider disabling cilium-health in ENI mode until we can confirm that it works reliably.