Use pitch positions for transition evaluations

[atas76]

Using just the match phase states for chaining transitions is a handy hack, however for both realism and accuracy we should be also using the pitch positions as reference. This was anyway the next step in the previous incarnation of this approach. One interesting thing to check is whether we have overfitting, with the low amount of seed data.