Just use Corosync over dedicated NICs, skill issue.
Kidding. Mostly. I actually have this exact problem and it's driving me insane.
Three-node Proxmox 8, all on Contabo dedis. Every 6-48 hours, one node gets STONITH'd by the others. No pattern. Logs say "node unreachable" then boom, fenced. VM migrations abort, HA restarts everything elsewhere, chaos.
IPMI shows all green. Network graphs look fine. I'm not running Ceph, just local ZFS.
Who's seen this? I will buy you a beverage.
oops: 0000 [#1] SMP