Add Boozer/SFL chart gate tests and signed chartmap Jacobian

[krystophny]

Follow-up to #397 (the intended base branch fix/395-boozer-cold-inverse was merged to main and deleted, so this stacks directly on main). Implements the ROADMAP straight-field-line/Boozer gauge gates as executable tests, ahead of the JOREK-sourced chart builder.

What

Signed Jacobian accessor. chartmap_metric_tensor returns sqrtg = sqrt(abs(det g)), which cannot distinguish a consistent chart orientation from a sign flip across a seam. New type-bound cs%jacobian_det(u) returns the signed det(dx/du) from the analytic covariant basis; metric_tensor is unchanged.

Gate (a) — signed periodic Jacobian positivity (test_chartmap_gate_jacobian): full-torus scan of a right-handed nfp = 3 rotating-frame fixture with probe pairs straddling every field-period seam, the theta seam, and the 2pi wrap; requires det > 0 everywhere (rho > 0), seam-pair continuity at solver-noise level, exact 2pi periodicity, and |jacobian_det| == sqrtg to 1e-9. A mirrored fixture must give det < 0 everywhere, proving the accessor is genuinely signed. Same positivity scan on the circular nfp = 1 fixture.

Gate (b) — bounded flux-label drift (test_chartmap_gate_fieldline): field lines of the production neo_circular_tokamak_field lie exactly on circular tori and are straight in the geometric angles; a tangency self-check verifies the closed-form trace against compute_bfield at 1e-12 first. 27 launches x 50 toroidal transits x 12 samples are pulled back through the warm invert_cart (orbit-pusher pattern, guess chained along the line). All 16 227 pullbacks must report LOCATED; max flux-label drift gate 1e-5 (measured 7.9e-7).

Gate (c) — launch-independent rotational transform (same test): endpoint transform estimate from unwrapped chart angles for the 3 radii x 3 poloidal x 3 toroidal launch ensemble; ensemble spread and deviation from the analytic transform (-B_pol/B_tor in the right-handed chart) gated at 5e-4, the analytic worst-case bound from the bounded periodic angle offsets (measured error ~9e-6).

Gate (d) — converged field reconstruction against the native evaluator (test_chartmap_gate_reconstruction): the consumer pipeline — native field sampled on chart nodes, periodic batch splines of (B_R, B_phi, B_Z, |B|), physical probe -> invert_cart -> spline evaluation — compared against the native evaluator at three resolutions (9,16,8) -> (17,32,16) -> (33,64,32). Gates: monotone error decrease, contraction >= 4 per refinement (measured 33x/15x vector, 7.7x/31x |B|), finest max relative error < 8e-5, the WP3 analytic |B| gate bound (measured 3.2e-8 vector, 5.3e-9 |B|).

Cleanup. chartmap_invert_cart computed a seed-refinement residual and error code (ierr_seed) that nothing read; chartmap_refine_cart_seed loses the two dead outputs.

Scope (honest limits)

Gate (d) runs against the analytic circular fixture, not JOREK: libneo has no JOREK-to-chartmap builder yet, so a JOREK-native gate (d) (against jorek_field on an AUG restart) must land together with that builder. The fixtures are smooth nested-surface charts; shaped/X-point geometry, non-nested regions, toroidal ripple, and the Fourier Boozer-transform pipeline are not covered. The constant-transform field makes gate (c) a launch-independence and absolute-value check, not a radial-profile-ordering check.

Tests

New: test_chartmap_gate_jacobian, test_chartmap_gate_fieldline, test_chartmap_gate_reconstruction (shared fixture writer chartmap_gate_fixtures, self-contained NetCDF in the style of test_chartmap_invert_cart). Full local ctest: 98/98 pass (Debug preset, gfortran), including all existing chartmap/Boozer inverse tests against the cleaned-up cold inverse.