The environmental variables that remain constant between the failures match what I used in my own test.
I also find it hard to believe that Bingos smoker VM can constantly change through OS's and submit PASSing test results for everything except OpenBSD occasinally if his VM indeed has bad memory. I suppose its possible, but it doesn't seem very likely.
It seems much more likely its some obscure bug in the Perl core, or with MakeMaker. If we knew what/why, it may be able to be worked around. But I will ask Bingos if he can do a memtest on his smoke machine and update.