Yes, exactly. If you look at the first image I posted, by the astronomers’ convention the angle from north to the point of maximum recession velocity is, by eye, about 30° so phi[i] should be around 0.5, and sure enough that’s where one out of 4 chains settles down to. But the algorithm doesn’t care about the convention and -0.5 works too provided the radial velocity flips sign as well (or in this model the coefficients of the polynomial representation). In the successful runs it appears that the parameters representing angles settle down to one or both modes before adaptation ends and the velocity coefficients likewise find the right mode(s).
Anyway I may have been too quick to dismiss using angles as parameters. My mistake was to declare them as bounded variables. What works instead is to leave them unbounded and, as you suggest, use a tight enough prior to keep them from hopping modes.
Thanks.