Single-contact CG comparison configuration

Native: 23.75 us. Our Model: 80.46875 us. Both: CG, discrete, 100 iterations, tolerance 1e-8.
The two *-options.xml files apply the benchmark overrides to the frozen options.
Include the original flex.xml inside every flex/flexcomp; set radius=1e-8 on each flex.
The sources directory preserves the original Newton/1-us base options and frozen build verification.
single-contact-config.json records all compiled mjOption values with solver/timestep overrides.
Contact override o_* options are inactive (override flag disabled); use the effective_contact entry.
Build: Linux x86_64, Python 3.12, MuJoCo 3.13.0 plus the supplied cap800 patch.
In the research repository: python3.12 experimental_cleanup/mujoco/build.py --jobs 4
Replay through experimental_cleanup/mujoco/run.py and the IROS_SHOWCASE CG cohort scripts.
A full result reproduction also requires the original 50-case reference/input batch, prepared meshes,
operators, material implementation and each study.json/result.json receipt on the research host. They are not bundled here.
Reported 42x / 56% / 42% headline values were supplied for this presentation, not recomputed by this exporter.
