The Physical Coding Sublayer is responsible for delimiting the frames, and sending them up to the MAC layer.
In Gigabit Ethernet, for example, the 8B/10B encoding scheme uses a 10 bit codegroup to encode an 8-bit byte. The extra two bits tell whether a byte is control information or data. Control information can be Configuration, Start_of_packet, End_of_packet, IDLE, Carrier_extend, Error_propagation.
That is how a NIC knows where a frame start and ends. This also means that the length of the frame is not known before it has fully decoded, analogous to a NULL-terminated string in C.
Once upon a time a twisted pair socket was only wired one way and the attached electronics couldn't change what each wire did. You were either a network device (hub/bridge/switch/router) or an end device. In order to electrically connect two network devices together you need a different cable than one used to connect an end device to a network device.
And thus the straight-through and cross-over cables were born.
Avoiding the usage of a second cable type (that would invariably lose its label and confuse the bejebers out of some network person months/years down the road when they pull it out of the bin) most devices intended to connect to both network devices and end devices had an uplink port that allowed the use of 'normal' cables.
It was as simple as that.
Edit: Google-Fu successful. It WAS ARCnet!
Why weren't switches/hubs designed from the beginning to use crossover cables instead?
Back when the 10base-T specification was still under consideration, the twisted pair architecture most common at the time in office networks was ARCnet. 10base-T wasn't ratified as an actual standard until 1990, later than I thought. Connecting ARCNet hubs together looks to have required a cable with flipped pairs from what ended up connecting to endpoint devices.
Since the standards committee would have been made up of veteran network engineers from the various hardware vendors and other interested parties, they had been dealing with the multiple cable problem for years and likely considered it status-quo. It is also possible that the 'draft' devices under development by the vendors also had electrical requirements for the cable, influenced as they were by ARCnet device manufacture. Clearly the committee didn't consider the use of multiple cable types to be enough of a problem to standardize the practice out of existence.
Best Answer
Doing some quick reading, it seems it is related to the Collision Detection part of CSMA/CD. If frames were too small on old broadcast media, then some collisions would be undetectable. Continuing my theme of automobile analogies today, it's for the same reason that we don't allow bicycles on high-speed highways - it's just not safe for them.