Should it be a requirement that ANY one-megabyte transaction that is valid
under the existing rules also be valid under new rules?

Pro:  There could be expensive-to-validate transactions created and given a
lockTime in the future stored somewhere safe. Their owners may have no
other way of spending the funds (they might have thrown away the private
keys), and changing validation rules to be more strict so that those
transactions are invalid would be an unacceptable confiscation of funds.

Con: It is extremely unlikely there are any such large, timelocked
transactions, because the Core code has had a clear policy for years that
100,000-byte transactions are "standard" and are relayed and mined, and
larger transactions are not. The requirement should be relaxed so that only
valid 100,000-byte transaction under old consensus rules must be valid
under new consensus rules (larger transactions may or may not be valid).


I had to wrestle with that question when I implemented BIP101/Bitcoin XT
when deciding on a limit for signature hashing (and decided the right
answer was to support any "non-attack"1MB transaction; see
https://bitcoincore.org/~gavin/ValidationSanity.pdf for more details).

--