Take the billow axis from the section pair, not a global projection

[1-Bart-1]

Problem

Billowing rotates each refined section's trailing edge about the leading-edge
span vector of the surrounding rib pair. The sign of that vector decides
whether the panel billows up or down, and it was chosen by projecting on
wing.spanwise_direction:

if dot(s_l.LE_point - s_r.LE_point, wing.spanwise_direction) >= 0
    le_neg, le_pos = s_r.LE_point, s_l.LE_point
else
    le_neg, le_pos = s_l.LE_point, s_r.LE_point
end

That test is well-conditioned only where the span vector is actually spanwise.
On a C-shaped kite the tip panels are nearly vertical, so their leading-edge
span vector is mostly x and z. Measured on the V3 kite's 37-section aero
geometry (spanwise_direction = [0, 1, 0], LE y running +4.167 → −4.167):

pair	|y_hat · spanwise_direction|
1 (tip)	0.033
2	0.071
3	0.115
4–33	0.23 – 1.00
34	0.115
35	0.071
36 (tip)	0.033

The outermost pair is separated by 11 mm of span. Since refine! re-runs on
the deformed sections each aero update, a flexing tip only has to move two
adjacent leading-edge nodes 11 mm relative to each other in y for that one
panel's billow to invert while its neighbours keep theirs.

Fix

Sections are guaranteed to run +y to -y (normalize_span_order! /
refine!'s sort_sections), so the pair itself fixes the axis sign with no
projection and no degeneracy:

diff_vec = s_l.LE_point - s_r.LE_point

spanwise_direction still sets the wing's overall handedness once, where it is
well-conditioned; it no longer decides each panel individually.

Effect

On a geometry whose tips are not near-vertical the two rules agree, so the
refined mesh is unchanged. On the V3 geometry above the old test also happened
to agree on all 36 pairs in the undeformed mesh — the change is a no-op there
— but it removes the 3 % margin the tip panels were relying on.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!