bug#24402: should-error doesn't catch all errors

[Gemini Lasswell]

Hello,

First, thank you for working on this bug!

I have a branch of Emacs which attempts to run all the tests under
Testcover, and I've tried it with Alex's latest patch. The patch has
fixed the problems with Testcover and should-error except for one weird
case. Here is a file that will let you reproduce the weird case on
master with the patch applied:

cyc-test.el
Description: Text document

This is an excerpt of test/src/fns-tests.el with a few lines added at
the end to invoke Testcover. It contains two nearly identical tests
which should both pass, but one passes and one fails. If you then edit
the file and comment out test-cycle-assoc (the one that passes) and run
the test again, then the failing test will pass. I've reproduced it both
in batch mode (without TEST_LOAD_EL=no) and interactively.