Hello Heng
This is a very good question. Looking at the process in more detail, when using Phase 3.
Initially, (and that is the key word) all spoke to spoke packets are switched across the hub. In order for a spoke to learn about the true NBMA IP address of another spoke, the NHRP redirect message is used.
So when a hub receives an IP packet inbound on its interface and switches it out of the same interface, it sends a special NHRP redirect message to the source indicating that this is a suboptimal path. It should look for a better way using NHRP resolution. The original packet however is still routed to its destination.
When the originating router receives the redirect message, it contains the destination IP address of the original IP packet as its payload. The router then sends the NHRP request for the redirected destination IP targeted originally, that is, the destination spoke. The resolution request travels via the regular IP routing path, through the hub until it reaches the target spoke.
The destination spoke responds to the resolution request using the IP of the source router sending it directly AND NOT THROUGH THE HUB, thus completing the spoke to spoke communication.
Now for your specific question, because the initial communication occurs via the hub, the hub still routes the packet successfully to the destination spoke. So the initial traffic sent does arrive at the destination spoke (via the hub) so it does not have to be resent. All subsequent communications that occur without the use of the hub will continue the flow of data.
I hope this has been helpful!
Laz