I'd go for a small server that returns a unix timestamp, the client fetches it, computes the clock drift (difference in time) and all the protocol related times are based on that drift. The clocks would still drift a little (they'd do it anyway on the system clock) and we would not get perfect synchronization (impossible in distributed systems), but it would solve our current problems.
The code is about 5 lines and some simple math (sums) when calculating the timestamps.
Why a server when so many are already out there?
I'd agree that we don't want bitcoin client bloat, and definitely NO WAY the bitcoin client should be setting the system time.
But maybe it is worth it in the error case for the bitcoin client to put a message in the log and attempt to use a network time instead of system time.
As for the servers out there, the bitcoin client already has http, right? Well, most http servers now provide the date in their headers. For example, in python:
import os, re, urllib
info = urllib.urlopen('http://www.yahoo.com/').info()
regx = r'Date:\s+[A-Z][a-z]{2}, (\d{1,2}) ([A-Z][a-z]{2}) (\d{1,4}) (\d\d:\d\d:\d\d)'
d, M, Y, T = re.search(regx,str(info)).groups()
m = 1+"JanFebMarAprMayJunJulAugSepOctNovDec".index(M)/3
print '%04d.%02d.%02d-%s' % (int(Y), m, int(d), T)
(that output format happens to work with "date -us" if you did want to set some system time)
of course it would be better to check a couple of sites rather than relying on just yahoo to keep their http server time set properly.
I'm just with 2 connections and 500 blocks.