Consolidate rzphi geometry extraction: share flux_surface_metric between FieldReconstruction and the metric/area helpers

[logan-nc]

Summary

Two independently-developed code paths derive flux-surface geometry (|∇ψ|, Jacobian, r²/η-offset and their θ-derivatives) from the same rzphi bicubic, via separate implementations. They should be consolidated onto a single rzphi-geometry primitive.

Background

• develop (src/PerturbedEquilibrium/FieldReconstruction.jl): _build_rzphi_geometry resamples the smooth rzphi primitives (r², η-offset, J, and their θ-derivatives) onto a coarse ψ grid and builds cubic interpolants for the PE field-reconstruction transform (added in PerturbedEquilibrium - PERFORMANCE - Coarse-grid rzphi geometry in field reconstruction #264, 3dcde8fd).
• feature/power-norm-eigenvalues (src/Equilibrium/FluxSurfaceMetrics.jl): flux_surface_metric is a pointwise evaluator of the same rzphi quantities returning (r, jac, |∇ψ|), shared by FFS (PowerNorm.jl) and PE (SingularCoupling.jl) after the 7dcfd5dc refactor.

These overlap conceptually: both pull r², the η-offset (deta), J, and the θ-derivatives off rzphi and combine them. They are not redundant code today (different consumers, different representations — real-space |∇ψ| weight + area integral vs. mode-space reconstruction geometry), so this is a maintainability cleanup, not a performance fix. The hot per-ODE-step path is already served by coarse ψ-splines on both sides (sqrtamat_spline in FFS, _build_*_interp in PE).

Proposal

Make flux_surface_metric (in the Equilibrium module) the single source of truth for extracting smooth geometry primitives from rzphi, and have both _build_rzphi_geometry and the area/weight helpers sample it. This keeps the GPEC gpout.f metric conventions in one place and prevents the two paths from drifting.

Risk / why this is its own issue

This crosses module boundaries (Equilibrium ↔ PerturbedEquilibrium) and carries real FMA-level numerical risk: the 7dcfd5dc refactor already documented benign 1-ULP shifts (≤ 2.2e-16) in 5 PE singular-coupling quantities purely from moving the metric block across a function boundary. Any consolidation must be validated against the regression harness (diiid_n1, solovev_*) with the usual 1e-8 noise thresholds, and the ψ-derivative handling needs care — _build_rzphi_geometry deliberately differentiates the resample spline rather than resampling the bicubic ψ-derivatives (those are only C¹ and ring near axis/edge), so a shared primitive must preserve that distinction.

Out of scope for

feature/power-norm-eigenvalues (PR #225) — that branch already centralized the FFS/PE |∇ψ| duplication; this is the follow-on consolidation with develop's reconstruction geometry.