This turned out to be not very serious after all. Phew!

What was happening was that those tests were pitching errors into the background error mechanism (because they arise at a point when we've gone through a C API that *can't* produce errors). Not sure why the valgrind runs were running `update` but the fix was to install our own logging `bgerror` and put the `update` inside the test. Then *all* builds see that *expected error* and the test will in fact fail without it.

It was mostly working as designed except for the error handling. No C changes needed.