I propose the following plan:

1) 0.9 ships with something dead simple, like a command to query what a node estimates and then clients just take the average, or cross-check a centralised estimate against the P2P network. It's fast to implement and simple, but not very secure or decentralised. However it will allow the feature to launch on some kind of reasonable timeframe.

2) We bump the protocol version and the tx message now gets an optional protobuf buffer stuck on the end. The first thing put in this protobuf is a list of the values of the inputs. Using this data, the fee paid by a transaction can be calculated. In step 2 the data is unauthenticated.

3) Some SPV wallets already set themselves up so that they sync with the network in the background, e.g. the Android wallet syncs at least every 24 hours. This should become more common, using scheduler capabilities built into most operating systems. When the wallet syncs with the network, it sets a deliberately very noisy Bloom filter on its peers and waits around for 30-60 seconds or so. The wallet observes some of the broadcasts taking place and records the hashes and associated fees that were paid to disk. Next time it syncs, it includes the observed hashes into the Bloom filter used to download the chain, and thus learns how quickly they confirmed. It can calculate its own fee estimate from that.

4) Finally, when we next hard fork, we make v2 transactions include the output value in the signature, same as the output script (this proposal has been on the forums for a while now). That allows the fee data added in step 2 to be cross-checked against the signatures on the inputs, thus authenticating it.

I think this is a small and easy set of steps that would make it quite hard to attack - malicious nodes could make it appear that some transactions never confirmed thus seeming to force the price up, but it's easy to simply exclude transactions which never confirm at all from the calculations. Plus of course you can cross-check nodes against each other to try and catch nodes that are failing to match transactions properly.

One obvious concern is what to do if nodes don't converge on very similar estimates. Wallets will always want to pay the lowest fee possible, so that means they'll always be riding the very edge of what's acceptable, opening up tx propagation to random flaky failures if fee estimates change whilst a transaction is in progress, or if some nodes don't calculate the same estimates as others.

If a wallet gets a reject message for a tx that has a fee that are by its own estimates acceptable, what should it do? What if only some nodes report that and others don't?