Independent check of a candidate plan against the receding I-STL constraints.
Used for two purposes with one definition:
* validation of every solver plan before it is installed (recompute robustness
independently of the solver), and
* the runtime check of a learned proposal: a proposal passes iff it is a feasible point of
the same synthesis problem, so accepting it keeps the recursive-feasibility argument
intact without solving.
Written separately from :mod:`rl_stl.synthesis.istl` (plain recursion over the ``stlpy``
tree, no Gurobi). ``rho_lower`` is the interval lower robustness of the enforced windows
with the semantics of the MILP: a planned predicate contributes
``a^T y_i - b - |a^T C| z(o+i)``; a constant history/root sample contributes ``+inf`` if it
meets its threshold and ``-inf`` otherwise. ``rho_lower >= delta`` (and bounds, dynamics and
the terminal block) is the acceptance condition.
An interval predicate ``[a]^T y - [b]`` (:class:`~rl_stl.stl.interval.IntervalPredicate`) contributes
the lower endpoint of its value over ``[a] x [b]`` and the sample box ``[C x_i - |C| z(o+i),
C x_i + |C| z(o+i)]`` (the I-STL lower bound, paper eq. (margin)); a constant sample is Boolean on the
lower endpoint with box radius 0 (measured, threshold 0) or ``|C| z(age)`` (cached plan, threshold
``delta``), as in the MILP.
:class:`CompiledWindows` evaluates the same semantics with numpy gathers compiled once per
``(problem, first_window)`` (a table of predicate values per sample, then one gather and one
min/max per tree level, as in :mod:`rl_stl.stl.jax_robustness`); :func:`interval_robustness`
is the plain recursion kept as the reference implementation (tests check they agree).