[plancherb1]

Summary

• get_tolerance refuses an unknown algorithm key instead of returning the default bucket, and the module refuses a per-robot override whose algorithm has no bucket. GRiD's CUDA equivalence modules had passed keys this table never had (forward_dynamics, f_ext_grad, f_ext_grad_so, fd) and compared under the 1e-7 default; gen3's continuous-joint bucket therefore never applied and its pinocchio leg was left out of the external-force checks.
• forward_dynamics and inverse_dynamics_gradient are algorithm-level buckets. The latter existed only as the gen3 override, so every other robot's gradient checks compared under the default with the same values. gen3 and h1_2 carry on forward dynamics the widening their inverse-dynamics buckets already have; no allowance changes for any existing check.
• The pinocchio oracle places the project's per-body wrenches through the mimic-inclusive scalar_joint_names, the list its f_ext gradient column already uses. Through the mimic-free list, h1_2's bodies after a mid-chain mimic joint received their neighbour's wrench and the mimic bodies none.
• tests/test_external_force_equivalence.py: inverse dynamics, forward dynamics, ABA and the inverse-dynamics gradients with a nonzero wrench on every body of every manifest robot, both bases, against pinocchio. tests/test_tolerances.py: the table is closed.

Verification

• Fast suite: 859 passed, 1 skipped (pre-existing singular-mass guard). Slow external-force cells (--runslow -k "f_ext or external_force"): 72 passed.
• Measured residuals with wrenches on every body (max over the dynamics samples, relative to output scale): gen3 1.2e-6 to 2.7e-5, h1_2 and fr3 at float64 round-off, every other robot at or below 1e-9.
• GRiD's three-way CUDA cells for gen3 (both bases), fr3 fixed and h1_2 floating pass on an RTX 5090 with this commit.

🤖 Generated with Claude Code