Note that this is a historic issue without any significance to modern network infrastructures. Ethernet networks are typically built with switches / bridges which create a single Ethernet collision domain for each connected station and typically operate in full-duplex mode, without ever employing CSMA/CD. Ethernet standards newer than 1000BaseX have dropped CSMA/CD and half duplex operation entirely.
But concerning the core of your question: there are a number of factors defined in the Ethernet standards preventing this from ever happening:
- the minimal frame length for Ethernet, which is 64 Bytes for everything up to 100BaseX, resulting in at least 512 bit times of transfer and is extended by padding to 512 Bytes for 1000BaseX (carrier extension)
- the maximum span for an Ethernet collision domain (2500m for 10Base5 and 205m for 100BaseT)
- the minimum nominal velocity of propagation for the signal in Ethernet cabling
- the maximum number of repeaters between every two Ethernet stations (4 with 10BaseX, 2 with 100BaseX)
- the maximum latency induced by a single repeater
Combined, this would ensure that a frame's transmission is never more than half finished before the frame start has reached the farthest edge of the Ethernet collision domain. The other half of the transmission time is being reserved for the collision signal to propagate back through the network, so collision detection is guaranteed to work under all circumstances.
Further reading: Charles E. Spurgeon, Ethernet: The Definitive Guide; Vijay Moorthy, Gigabit Ethernet