Fix inverted q95 to qbar transform in Peng TART scalings

[dallonby]

Closes #4532

Overview

The q95 → qbar transform in the two i_plasma_current=2 (Peng TART/STAR) call sites multiplied by 1.3*(1-eps)^0.6 where inverting the documented relation q95 = 1.3*qbar*(1-eps)^0.6 requires division. Detail and pre-#3320 history in the issue.

Changes

• process/models/physics/plasma_current.py (calculate_plasma_current_peng) and process/models/physics/plasma_fields.py (calculate_surface_averaged_poloidal_field, TART branch): divide by 1.3e0 * (1 - eps)**0.6e0 and document the inversion in a comment.
• New tests/unit/models/physics/test_plasma_current.py (first unit tests for this module): hand-derived reference values for the corrected current and bpol at ST parameters, a round-trip test on the transform (including qbar > q95, which the forward relation requires for any eps > 0), and a consistency test that both call sites share the same qbar.
• tests/unit/models/physics/test_physics.py: the four existing golden values for calculate_plasma_current_peng / calculate_surface_averaged_poloidal_field are updated to the corrected values. Justification per CONTRIBUTING: these goldens pinned the inverted transform; both sit at conventional aspect ratio (A=2.7/3.0) where the correction is only −3.0%/+3.9%, which is why they never exposed the bug. Comments in the diff give the scale factor (1.3*(1-eps)^0.6)^2.

Behavioural impact

Only i_plasma_current=2 runs change. No shipped regression input uses it, so regression MFILEs should be unchanged. External ST studies using the Peng scaling will see plasma current drop ~35–60% at A=1.4–1.8 — that is the fix, not a regression: with it, icurr=2 agrees to ~2% with the independent icurr=9 scaling on the shipped ST geometry (22.5 vs 22.9 MA), where it previously disagreed by ~55%.

Verification

• Full unit suite passes locally (846 passed / 4 skipped, Python 3.12).
• If maintainers know of a historical STAR-code benchmark for icurr=2, that would be the ideal additional cross-check.

Found during an independent audit of v3.4.2.

🤖 Generated with Claude Code

[dallonby]

Correction to the 'Behavioural impact' section above: the regression MFILEs are not fully unchanged — the diagnostic comparison entry c_plasma_peng_double_null (the Peng-scaling current reported alongside every run's chosen scaling) changes by the fix's factor even on inputs that use other i_plasma_current models. Verified on st_regression: every design variable is identical to the last digit; only this one diagnostic row moves (3.291e7 → 2.102e7 A, −36%). So the 0.2%/5% regression jobs will flag exactly that one variable per tokamak input. All actual design outputs are unaffected.