The two codes you imported that had that two-bit repeat burst were taken correctly and repeated because they have that part. The two without that were taken incorrectly and transformed from signals that shouldn't have repeated to signals that did.

That still may leave some variety in exactly what is repeated and how fast, explaining any differences you see in how the device responds to the four repeating signals.

Or were you asking from the point of view of the device rather than the remote? All NEC1 signals have that repeat part sent by the remote. The device is free to ignore that repeat part if repeating is a bad idea for that particular function. That decision is in the device, not in the remote (for other protocols that decision might be in the remote).
.