If it's evaluated first, it doesn't really get you the behavior it was created for. The behavior I described above, the behavior that is documented, and the behavior @hubitrep is seeing in 2.3.9 are all (the same and) correct.
Specifically, see the example in the doc that triggers on mode change. This could never work any other way. It's really only a concern when the trigger and required expression refer to the same event. It's the value at the moment just before triggering (or at but before the resulting state is the case) that matters.
Looking into possible changes in 2.4.0...
