I confirm that casting the array as pm.Data() also solves the issue for me. This solves the issue for various combinations of homogeneous tails/heads datasets. I agree with your guess, the code likely interprets the full array as a single observation np.array([0]) instead of the full length, hence the Beta(2,1).
Thanks to you and to @ricardoV94 for the fast reaction. And thanks for opening and issue on the Github repo.