I found that a pair of VMs I thought I had that was redundant had quietly ended up on the same Proxmox host. The services were redundant at the application layer, but not at the physical failure-domain layer on my Proxmox hosts.

I wrote up what I found, how VM placement can drift during migrations and maintenance, and how Proxmox HA resource affinity rules can help keep redundant workloads on separate nodes.

The bigger lesson for me was that two copies of a service do not necessarily mean two separate failure domains.