I think direct feedback in PM is possible, at least for low modulation indices. The Yamaha implementation is in software, so a one sample delay is the easiest implementation (otherwise you'd have to solve a rather complex equation at every step).
The distinction that I’m making is exactly that. You can’t incorporate direct feedback into a phase modulation synthesizer, you need to incorporate delay into the feedback path. By comparison, you can use direct feedback with subtractive or FM synthesis.
The one-sample delay basically turns the feedback into an iterated function system. With certain parameters, you end up with aperiodic results—cool in theory, but in practice, often just used as a noise source.
But if we take the simplest example, f(x) = sin(x + c*f(x)), we can solve it, (numerically). If I run it up to 0.5π for c=0.5, I get something that's (obviously) very close to the function with a short delay. So I take that as the possibility of direct feedback, although in a electronic circuit, feedback also has a minute delay.
