benchmarks.hota.learned_modules

The ``learned`` tracker (``build_learned_tracker`` in ``tracker.py``) replaces the closed-form ``Identity``/``Replace`` embedding filter with a pair of small learned modules wrapped by ``LearnedProcess`` / ``LearnedObservation``: - :class:`Propagator` is the predict step — a residual MLP that nudges a track embedding forward in time and renormalizes it onto the unit sphere. - :class:`Fuser` is the update step — a gated residual fuse of a track embedding with its matched detection's embedding. Both are autograd-native, so the same cosine/Sinkhorn association objective used at inference can train them (see ``train_learned.py``).