velocity_3D_trailing_vortex_semiinfinite! returns NaN on its own axis at its start point, and its core is not linear in the distance #399

Found while fixing the bound-segment kernel in the #318 pull request, which moves velocity_3D_vortex_segment! onto a closed-form core that is linear in the distance to the axis and differentiable through it.

velocity_3D_trailing_vortex_semiinfinite! (src/filament.jl) still has the old shape, with two differences from the bound kernel:

• Its Lamb–Oseen core epsilon grows from zero at x1. At XVP == x1 (or on the axis within rounding of it) both the outside test and the < 1e-12 * epsilon guard are false, the core branch normalises a zero radial vector, and the result is NaN.
• Inside the core it returns the core-boundary velocity unscaled, so the velocity is not linear in the distance to the axis: it jumps in direction across the axis, and ForwardDiff sees zero there.

No evaluation point the solver builds today sits on a semi-infinite trailing axis (aero centres and control points are mid-panel, and trailing filaments start at panel edges), so nothing measurable moves. The question is whether the inside-core model should be the linear (Rankine-like) one the bound and finite trailing kernels use, which would change results for any point inside a wake core.

What I would do: scale the core branch by distance / epsilon as the bound kernel does, give it the same closed form, and guard the zero-core case the same way. That is a model change, so it is yours to decide.