If it's internally stored, it places a significant burden on manufacturing, as they have to track which addresses have been used and ensure they're NEVER used again.
no "significant burden", just put a DS2411 on the board and let Maxim do it for you at a cost of $0.33

I agree, though, that if you decide to use that scheme, it IS simpler for the end-user.
who cares about 'simpler' as Andy said If users can fiddle with it, then they will mess it up sooner or later!

It increases the initialization time from several seconds to many hours, though, since it takes a while for the master to poll 2^64 addresses, waiting for all responses, then recording and tabulating the non-nil ones.
nope, it's quite fast. With the proper algorithm you can 'find', say, 64 slaves (each with a 48 bit address) in about 2 seconds. Then assign them an 8 bit 'node address'

Clearly that would force some more complex protocol, with the baggage that brings with it, as polling is, under such conditions, far too time-consuming.

the remainder is based on 48 bit node address, max 250 nodes (with random 48 bit addresses changed by a boot routine to 1 byte)
Once you have your "node list" it does not need to take more than the time for 2 bytes out and 1 byte in to poll.

Erik