As a followup to this thread, I concocted a dirty test which aims at revealing the paged nature of EEPROM in AVRs.
It writes an alternating pattern of 55/AA into a certain EEPROM position in batches of 1000, and then reads back this position to find out wear and tear. It also writes this alternating pattern once per batch into neighbouring bytes and reads it back at the end of batch - that should reveal whether any paging is involved in rewriting the EEPROM. Results are printed through UART.
I am now running it on a "spare" ATM128. Currently, it runs batch #840, no, wait, #841, so it is well over the stated endurance limit, and it's still OK - but that's no surprise, nominal endurance certainly guarantees proper readback at temperature extremes etc. which I don't exercise, of course. One "batch" lasts around 9s (as one EEPROM write is nominally 8.5ms at the 'M128), so 1M writes happens in around 3 hours. I expect the first errors will occur within 10M writes, i.e. a couple of days' "work".
Other experimenters willing to kill an AVR of choice are welcome to try to adopt this snippet, of course.
PS. #871 and counting...