t1: Server sends message (1) to Client B
t2: Client B fails to decode message (1) due to an outdated key. Sends failure notification to Server
t3: Server sends rekey notification to Client A
t4: (usually <400ms or >1 second) Client A sends rekeyed message (1') to Server for Client B
For t4, if the Server notices the response from Client A came faster than a user could ever have seen and responded to a notification, it now knows Client A is more susceptible to a MitM attack. If, on the other hand, Client A takes an appreciable amount of time to send the rekeyed message (say, 1 second or more) then we know that either there was a network latency, or the user had to respond to a rekey notification before the message was resent. We should be more careful about launching a MitM attack on this class of user.
So, without changing the actual _actions_ that the Server sees, but just the pattern of the timing of those actions, we're still leaking important information.
Note: all numbers completely made up.