Compile ``stlpy`` formulas into jittable JAX robustness functions.
``stlpy`` owns formula construction; this module only evaluates the standard
discrete-time quantitative semantics (Donze and Maler) that ``stlpy``'s
``robustness()`` implements, so the same formula object can be scored inside a
JIT-compiled, vmapped MJX rollout:
- ``LinearPredicate``: ``rho = a^T y_t - b``
- ``STLTree`` "and": ``min_i rho(child_i, t + timesteps_i)``
- ``STLTree`` "or": ``max_i rho(child_i, t + timesteps_i)``
The tree is walked **once in Python** at compile time. Children that are the same
object (``always``/``eventually`` repeat one subformula over a time interval) are
grouped, their time indices are deduplicated, and each group becomes a single
gather. The resulting function has only static shapes and indices.
Signals are ``(T+1, d)`` arrays (time first), unlike ``stlpy``'s ``(d, T+1)``.
``NonlinearPredicate`` is not supported (the MICP layer does not support it either).
Reward timing is *not* decided here; see :func:`stl_hinge_penalty` and the X2
environment docstring for how windowed robustness is turned into per-step rewards.