I've had many things like that over the years. Sometimes a simple "trace" with the output looking for failed opens/reads/writes would disclose the culprit: it amazed me how many programs didn't bother to check such things because the file "had to be there" or that the perms would always allow their reads or writes.

But I've also had more than a few where I guessed that the problem was flaky hardware and had to write code that exercised it enough to prove the fault.