Improve `defer-failures`

[sellout]

defer-failures now preserves the original error message, and we can check that message in test cases. There is no longer any need to disable defer-failures (or jump through the other hoops previously listed in the README) in order to debug test failures.

This also enables us to address #45.

[sellout]

~The formatter failures are annoying – despite them being in Nix, I can’t replicate them locally on the same architectures, and the same checks are passing on aarch64-*.~

Nevermind – apparently this was a transient failure.