[krystophny]

Risk tier

• T3: physics, output behavior, coordinate convention

Correctness contract

Intended behavior change

On a wall chart that reaches past the wall, the radial Meiss ODE $\lambda' = -h_r/h_\varphi$ has no solution where $h_\varphi = \mathbf b\cdot\mathbf e_\varphi$ changes sign. On the W7-X KJM map2disc chart (LCFS + 0.8 m) in the frozen-current field of W6 t = 1.05 s this happens on 1588 of 4032 radial lines between the wall and the chart edge, and init_meiss stopped with "Meiss transformation ODE failed".

With wall_input set, a line now ends at the last node where $|h_\varphi| \ge 0.1,|h_\varphi(r_\text{min})|$; $\lambda$ and $\chi$ are held from there (cut_line). Along the W7-X Meiss lines the smallest ratio met before the wall is 0.21, except on one line that runs through an opening of the STL wall and never meets it. A marker beyond the end of a line of its grid cell therefore stops as a numerical domain exit (orbit_exit_code 101) instead of being traced in a non-canonical field (beyond_meiss_cut, checked after the wall query of each symplectic microstep, also for Albert via its Meiss radius). Initialisation reports how many ended lines meet the wall first.

The line through the node where $A$ is gauged to zero starts with $\chi = \chi' = 0$; with a purely relative error scale its first step fell to roundoff (W7-X: ierr = 1 at node 2 of $\theta = 0$, $\varphi = 0$ with a smooth right-hand side). The ODE call now passes libneo's scale_floor, set to 1e-6 of the change of $\lambda$ and $\chi$ over one grid interval, so it prevents the underflow without relaxing the requested tolerance (test_spectre_freeboundary still sees its step-limit failure).

Behavior that must not change

Without wall_input no line is cut and an ODE failure still stops initialisation.

Coordinate / unit conventions

Unchanged.

Numerical invariants

The canonical form holds up to r_cut(i_th, i_phi); no marker is traced beyond it.

Tests added

• unit: test_meiss_line_cut (chart whose outer surfaces tilt toroidally, $h_\varphi \propto B_0R_0 - B_z d\rho^6\sin\varphi$): some lines end, none before $r = 0.85$, the canonical field stays finite on them, and a point past an end lies outside the domain. It does not compile without the change.

Golden-record impact

• unchanged
• changed: the floor can change Meiss and Albert transformations at the level of the ODE tolerance; the merge queue's golden records decide.

Verification

The Meiss tests (test_meiss_line_cut, test_field_can_meiss_config, test_chartmap_wall_losses, test_profile_radius_meiss*) pass locally with GCC 16. The W7-X numbers come from diagnostics on sCluster in the W6 t = 1.05 s frozen-current field.

Chris&AI