I note that Trac #4450 had exactly the same diagnosis that we've just rediscovered for Trac #8611, 8 years ago. So to avoid this happening again for more benchmarks, why don't we just do +RTS -V0 for all the single-threaded benchmarks in nofib? That is, add it by default to all parts of nofib except nofib/smp and nofib/parallel.
I was thinking that +RTS -V0 was just a temporary workaround until we have a story for Trac #16065. But if we want to allow nondeterministic allocations in regular program execution then I'll close Trac #16065 as invalid and will prepare a patch that does +RTS -V0 for all single-threaded benchmarks.