I have written a reference implementation and BIP draft for a soft-fork
change to the consensus-enforced behaviour of sequence numbers for the
purpose of supporting transaction replacement via per-input relative
lock-times. This proposal was previously discussed on the mailing list in
the following thread:

http://sourceforge.net/p/bitcoin/mailman/message/34146752/

In short summary, this proposal seeks to enable safe transaction
replacement by re-purposing the nSequence field of a transaction input to
be a consensus-enforced relative lock-time.

The advantages of this approach is that it makes use of the full range of
the 32-bit sequence number which until now has rarely been used for
anything other than a boolean control over absolute nLockTime, and it does
so in a way that is semantically compatible with the originally envisioned
use of sequence numbers for fast mempool transaction replacement.

The disadvantages are that external constraints often prevent the full
range of sequence numbers from being used when interpreted as a relative
lock-time, and re-purposing nSequence as a relative lock-time precludes its
use in other contexts. The latter point has been partially addressed by
having the relative lock-time semantics be enforced only if the
most-significant bit of nSequence is set. This preserves 31 bits for
alternative use when relative lock-times are not required.

The BIP draft can be found at the following gist:

https://gist.github.com/maaku/be15629fe64618b14f5a

The reference implementation is available at the following git repository:

https://github.com/maaku/bitcoin/tree/sequencenumbers

I request that the BIP editor please assign a BIP number for this work.

Sincerely,
Mark Friedenbach