I should mention that a high level of test parallelism on the part of my worker might have be a contributing factor in this most recent case.

The worker was recently upgraded to a faster 4-core VM, but with limited I/O.  In a test run the test processes invariably end up stuck on I/O heavy tests, idling the CPUs.

So I've been running the tests under -j8, as I found it the most effective combination of supporting tests stuck on I/O while keeping the CPUs busy, but it does mean that in some cases there's a lot pending on the CPUs, and depending on the exact test ordering in a run presumably some more sensitive tests could be impacted.

I have in fact seen an increase in random tests generating warnings (fail, then pass) than the worker had previously.  I suspect the benefits of the extra parallelism on total test time (-j8 is about 20% faster than -j4) probably isn't valuable enough and will most likely be reducing it a bit.