expect_failure() should report that the correct number of failures was indeed observed

```r
expect_failure({
  pass()
  fail('x')
})
# Error: Expected exactly one failure and no successes.
# Actually succeeded 1 times
```

It's not clear from the message itself that the failure side of the equation is balanced -- I was reading through logs and puzzling over why the test didn't fail as I was finding it to do interactively.