Gravity-control implementation

The implementation uses the shared library. The study plan defines the question, finite allowance and gates. A software pass alone does not qualify an observed-data fit.

1. Native interface and derivatives. Add explicit height-gravity entry points to analysis/lll/ilvis0_native.c and analysis/lll/ilvis0_tangent.cpp. Preserve existing entry points and private constant-step callers. Extend ilvis0_exploratory.load_kernel and ilvis0_shape.load_tangent with optional gravity_anchor_height_m; None keeps the original interface and metadata. The height adapter binds a fixed anchor while preserving the call signature expected by training and forecast code. Evaluate gravity from predicted midpoint height, including automatic derivatives through that state.
2. Focused qualification. Write failing tests first in analysis/tests/test_ilvis0_height_gravity.py. Compile historical source from the verified ESA pilot archive for constant-mode comparisons. Test analytic vertical motion, invalid anchors, value/tangent agreement and all 27 derivative columns for all three candidates. Retain both processing cases. Run through analysis/tests/run_numerical.py and include affected native/compensation regression checks.
3. Independent generated study. Extend the existing independent Cartesian truth generator with explicit climbing/descending profiles and a separately implemented gravity law. Keep its old defaults unchanged. Implement a small lll.studies definition/evaluator and saved-only reporter calling shared rate, calibration and motion components. Bind all source, environment, construction and finite native-work limits before running the 48-case matrix. Publish every failure and verify saved positions, derivatives and accounting without replay.

The independent truth must not use the predictor's gravity helper or equations to generate the expected trajectory. Later receiver positions cannot enter the height-gravity forecast. Changing the source requires a new freeze; completed allowances and historical dependencies remain unchanged. Archive writes are serialized.