Every generation with a reference is an argument between the picture and the words. Reference strength decides who wins, and the default is rarely the right answer for production work.
High values hold identity and fight the prompt: ask for a different angle and you may get the reference angle back. Low values follow the prompt and let the subject drift. Which failure you prefer depends entirely on whether identity or staging is the thing that must not move.
For a recurring face, bias high and change the staging with a pose or depth condition rather than with words. That way the two controls are not competing for the same job.
Why does my subject refuse to change angle?
Reference strength is probably too high, so the reference is overriding the prompt. Lower it and control the angle with a pose or depth condition instead of with words.
What is the right setting for a recurring presenter?
High enough to hold the face, with staging controlled by a separate conditioning signal. Asking one dial to do identity and composition at once is why it feels like it never obeys.