[cmccomb]

Summary

• Recompute full 225 s numerical sensitivity from checked MATLAB traces and independent Python physical runs driven only by pinned HDF5 and sea phases.
• Preserve the unchanged pointwise gate as a strict expected failure. The MATLAB 0.125/0.0625 ms source pair differs by 0.010800 rad pitch; independent Python 1/0.5 ms runs differ by 0.005127 rad, both above the 0.0005 rad gate. All five original pointwise measures fail both step comparisons.
• Audit the fixed-input RTI QP on checked source states, forecasts, and preceding plans. Its Hessian is well conditioned, Python bound violations are zero, and resetting the preceding plan changes the actual optimization problem rather than just solver accuracy.

Verification

• pytest -q tests/test_wavestar_nmpc_stability_diagnostic.py tests/test_wavestar_nmpc_qp_recovery.py tests/test_wavestar_nmpc_controller_portable.py: 3 passed, 1 expected failure.
• tools/wavestar_nmpc_autonomous_step_sensitivity.py completed both independent 225 s runs on macOS ARM64 with NumPy 2.5.4 and SciPy 1.18.1. Checked output and QP reports are committed.
• git diff --check passed.

Scope

Diagnostic only for #623. No production controller, original acceptance limits, case status, or badges change. The issue stays open until a numerically stable 225 s source target and independent physical Python run satisfy the original limits on supported runners.