KineticForces — dW_kinetic (imaginary NTV energy) ~17–19% off Fortran on DIIID benchmark

[logan-nc]

Summary

On the DIIID PENTRC benchmark (PR #112, "KineticForces Module"), the NTV torque metrics clear the 5 % acceptance bar against Fortran GPEC, but the imaginary energy-conversion-rate metrics — dW_kinetic — do not. They drift ~17–19 % from Fortran. This was flagged during PR review and explicitly accepted as a non-blocking follow-up; this issue tracks resolving it.

Evidence

Torque (headline result) — passing:

Metric	Julia	Fortran	Error	Status
T_total_fgar (N·m)	1.805189	1.780406	1.39 %	✅ < 5 %
T_total_tgar (N·m)	0.991272	0.948962	4.46 %	✅ < 5 %

Imaginary energy-conversion-rate dW_kinetic — out of band:

Metric	Julia	Fortran	Error	Status
dW_total_fgar	0.160012	0.136829	16.94 %	⚠️
dW_total_tgar	0.133295	0.112297	18.70 %	⚠️

This is a regression from PR #224's pre-merge values (dW_fgar = 2.68 %, dW_tgar = 4.81 %). An A/B test on merged HEAD with the pre-port energy integrator swapped back in gives identical numbers, so the energy-integrator port itself is ruled out as the cause.

Leading hypothesis

Develop commit 083b0cc0 ("EQUIL - BUG FIX - Snap periodic theta endpoint at 2D interpolant construction sites") changed how 2D equilibrium interpolants evaluate near the periodic-θ boundary. Bounce-data computation reads those interpolants, and the imaginary part of the kinetic resonance integral is far more sensitive to that boundary than the real (torque) part — consistent with T passing while dW shifts by ~12–13 % from #224's pre-merge value.

Proposed next steps

1. A/B the bounce-data interpolant reads with the pre-083b0cc0 θ-endpoint handling to confirm or deny the hypothesis.
2. Inspect the imaginary part of the kinetic resonance integral near the periodic-θ boundary for sensitivity to endpoint treatment.
3. If confirmed, decide whether the fix belongs in the interpolant boundary handling or in how compute_bounce_data samples θ.

Reproducer

BENCHMARK_TGAR=1 julia --project=benchmarks/.benchmarks_env benchmarks/benchmark_diiid_kinetic.jl

References

• PR KineticForces Module (formerly PENTRC) #112 — validation comment, sections 2 (NTV torque benchmark) and 5 (dW disclosure).
• Branch feature/kinetic-validation-proofs — raw (ψ, dT/dψ, T) arrays under benchmarks/kinetic_validation/kf_{fgar,tgar}_profiles.dat.
• Suspected cause: commit 083b0cc0.