As a follow-up, I mostly left the script alone, since the wife is in the final weeks of a really big analytic paper and I can't risk slowing her down with my debugging. But I did double the memory on the machine (to 12GB), and saw the clock-time cost of the Fisher-Yates shuffle drop from the ~hour mentioned above to about 45 seconds to do all 43-million iterations.

So I think we can safely say it was a swapping issue...

Thanks for all the kind suggestions and analysis!