reg_field: membrane penalty runs at half strength inside make_kernel_bending_rls

[balbasty]

Summary

make_kernel_bending_rls (regularisers/field/2d.h, and the 3D analogue) fills the kernel table via make_kernel_bending(kernel, ..., nc) and then uniformly rescales kernel[1..5] by 0.25. That rescale is correct for the bending part of the coefficients, but kernel[1] = -4·b·vx·(vx+vy) - m·vx (and the vy analogue) also mixes in a membrane term, which wants a 0.5 scale factor (matching the plain membrane_rls kernel), not 0.25.

Symptom

With a uniform weight map w≡1:

• bending_rls reproduces matvec_bending exactly for absolute-only and bending-only inputs (diff ~1e-16/1e-14).
• For membrane-only input, the diff is ~1.07 (not the ~1e-16 you'd expect from an exact reduction) — the membrane contribution is running at the wrong magnitude specifically when computed via the bending-order kernel path (i.e. when bending != 0 alongside membrane != 0, or more precisely whenever make_kernel_bending_rls is the constructor used).

This is a magnitude/scaling bug, not a self-adjointness bug — the operator stays symmetric either way, it's just computing the wrong coefficient value for the membrane term when bending order is active with RLS/JRLS weighting.

Suggested fix

Split the two scale factors explicitly instead of a single uniform 0.25:

kernel[1] = -b * vx * (vx + vy) - 0.5 * m * vx; // was: -4*b*vx*(vx+vy) - m*vx, then *0.25
kernel[2] = -b * vy * (vx + vy) - 0.5 * m * vy;
kernel[3] = 0.25 * b * vx * vx;
kernel[4] = 0.25 * b * vy * vy;
kernel[5] = 0.5 * b * vx * vy;

(matching the existing pre-scale-then-multiply structure, just correcting which factor applies to which term). Verify: keeps the self-adjointness fix from #38 intact, and makes w≡1 bending+membrane+absolute reduce exactly to the unweighted matvec_bending/matvec_all at that combined order.

Related

Also worth a deliberate decision (separate follow-up, not blocking this fix): bending_rls's kernel[0] (the absolute term) is left unweighted by the RLS weight map, whereas membrane_rls's kernel does modulate absolute by the weight — an inconsistency between the two orders' semantics that predates this issue.

Surfaced during the investigation for #34 (bending RLS/JRLS self-adjointness), not folded into that fix since it's an independent magnitude issue.

Workstream: claude-jitfields-to-fastfields