If the client can usually work in streaming, then clearly none of the suggested possible constraints applies.
I confirm that the described pattern corresponds to the Stream-sense mechanism, which opens a polling session when the first bind_session yields no answer to the application for more than 5 seconds.

Now, the most likely explaination is that there are delays either in the client machine or in the network.
In fact, the request times reported in your logs are quite separated; by the way, do you know if the server and client clocks were in sync when you took the logs?
Can you see the exact times of the single outbound and inbound packets of the requests to the Server on the client machine? They could help discriminating where the delays are produced.