The last 4 bytes of the minimum frame is the CRC/FCS field. So, if the station is transmitting the CRC/FCS field (Bytes 61 - 64) and receives a collision, it should retransmit the frame at a Data Link layer, with only a minimal delay. This would be a fragment, a collision after we have started transmitting data. As opposed to a preamble collision.
If the last bit of the 64th byte is transmitted by sender and the collision occurs within a hub along the path, the sender would not know. As a result, the upper layer would have to wait the retranmission timeout before resending the frame.
Here is one way I figure out which one it is.
If the Data Link layer retransmits, both the collision frame and the retransmitted frame will have the same IP identification number in the IP header. If the sender didn't know the frame suffered a collision, the retransmitted frame will have a different IP identification number.
You are correct in your statement. If the collision occurs during the last 4 bytes of a 64 byte frame and the sending host detects the collision, it is not a late collision.
This is why I am thankful for switches and full duplex, so we don't have to deal with these stinking collisions!
Sorry about the long winded answers....
Mike