If the recent problem reports from Kenoyer and Eustace are the same
thing you saw, then the problem must exist in 7.1 as well as current
sources.
The best idea I've had so far is that there's something going wrong in
old-style VACUUM, probably in the chain-moving code which is horridly
complex and has had several bugs found before. (Eustace's
heavily-updated single-row table would present a very long chain, so
that seems to fit.) But this does not explain your problem seen with
7.2 sources, unless you were using VACUUM FULL. Were you?
regards, tom lane