Job performance can be defined many different ways, and the most common in field studies is what is most common in the workplace: supervisory, and possibly peer, evaluations. It's not the perfect criteria, but the "perfect criteria" for many jobs exists only in the theoretical. When appropriate, "work performance" is also defined as sales numbers, widgets produced, time-to-complete, etc. It just depends on the study.

> So the test has to be matched to a position.

In many cases you're right (e.g., work samples), but not always. g-centric tests have demonstrated high validity across jobs and contexts. Tests for cultural fit (in which the desired outcome is typically workplace satisfaction and tenure) are typically validated across a wide range of jobs, and so they are applied across those jobs. Things like integrity tests and personality tests can have meaningful validity across jobs, even without seemingly job-relevant content.