Citrix recommended practices for NetScaler high availability (HA) warn against cross-over cables for HA synchronization, as it will result in a bridging loop. This behavior was noted on a customer site with this configuration when testing failover with no datacenter access or network configuration knowledge. While rebooting the secondary, the primary would lose its status and demote itself to secondary. When the secondary came back up, it would make itself the primary.
While the secondary was being rebooted, the primary was receiving the signal of being partially down, and the HA state becomes inactive. The primary will flap between partial and complete fail and will be the secondary until the secondary comes back up. Whichever NetScaler boots first will be the primary, so the secondary then becomes the primary and then sync will resume.
A good way to test if this is the case is to turn off HA Monitoring, while leaving the HA Heartbeat on. The failover behavior should return to normal with Monitoring off.
***Please note: Neither direct cabling nor turning off monitoring is recommended.