Systemd socket activation application support limits for zero-downtime upgrades

Systemd socket activation is frequently discussed as a path to zero-downtime application upgrades on Rocky Linux because the listening socket can stay open while the service binary is replaced. The documented requirement is that the application must accept the file descriptor from systemd rather than binding the port itself.

For legacy binaries that do not support inherited file descriptors, administrators must introduce a wrapper or modify the application. The unresolved behavior concerns the transition itself: when the old service process is stopped and a new process begins accepting from the same socket, it is unclear whether systemd queues incoming connections during the gap and whether established connections to the old process are terminated immediately or allowed to drain.

What are the precise support limits for systemd socket activation when a small application is migrated without dropping active connections?