I still think using the version and timestamp fields in the block header
are simplest and best.

Advantages:
  Available to SPV nodes with no change to the network protocol
  Available after headers downloaded, before full block data is available
  Once well past a fork, allows all block validation except validation
against the UTXO to happen in parallel, out-of-order, independent of any
other block.

Disadvantages:
  Not monotonically increasing


I think discussion about transactions in the memory pool are just a
distraction: no matter what criteria is used (timestamp, height, median
time), a blockchain re-organization could mean the validity of transactions
you've accepted into the memory pool (if you're accepting transactions that
switch from valid to invalid at the consensus change -- Core tries hard not
to do that via IsStandard policy) must be re-evaluated.

I don't strongly care if median time or block timestamp is used, I think
either will work. I don't like height, there are too many cases where the
time is known but the block height isn't (see, for example, the
max-outputs-in-a-transaction sanity check computation at line 190 of
bitcoin-tx.cpp -- bitcoin-tx.cpp has no idea what the current block height
is).


-- 
--
Gavin Andresen