material: torch twin of the bicycle material rollout; parity with libcvc; trainer on the bicycle

[transfix]

The Python side of libcvc P8b (transfix/libcvc#642): training the material model (CoefEnergyNetMaterial) through the deployed kinematic bicycle instead of the point-mass surrogate.

What

• grl_snam/material_bike.py: integrate_bicycle_material in torch, with the material surrogate's force law driven through drive.cpp's controller:

  • pure pursuit plus the (F_rep + F_mat) steering bias;
  • the corner, stop, creep and reverse governors;
  • d_cap and the kinematic update.

  Every branch is a torch.where, so autograd follows the branch each step took, as libcvc's hand-written adjoint does. Guarded sqrt/norms keep an unselected zero from producing NaN gradients.

• nav_native: integrate_bicycle_material / integrate_bicycle_material_vjp over the new pycvc entry points, behind HAS_BICYCLE_MATERIAL_ROLLOUT.

• MaterialTrainer(bicycle=BicycleParams(...)): takes the loss through the bicycle, behind HAS_MATERIAL_TRAINER_BICYCLE; .integrator reports which rollout is in use. The module still imports without torch.

Tests

• tests/test_material_bike_parity.py:
  • forward parity (rtol 1e-4) and gradient parity (torch autograd vs the C++ adjoint, rtol 1e-3) on two batches;
  • a pure-torch check that the twin drives to its goal within its speed limit.
• tests/test_material_train_native.py: the trainer descends on the bicycle, and a zero wheelbase is refused.

Measured locally against a pycvc built from transfix/libcvc#642: forward within 5e-7 relative and gradient within 3e-6. Two independent derivations (hand-written reverse mode and autograd) agree to float precision. The GRL-SNAM material suites pass (42 passed; the 4 skips are a missing CUDA device and the pure-mode tests).

The native tests skip in CI until a pycvc with the bindings is installed; the pure-torch test runs now.