180° heading flip at `init!`/`reinit!`

[ufechner7]

• 180° heading flip at init!/reinit! for wings whose CAD-frame position
  sits only millimeters off the transform's rotation axis (e.g. the
  ram_air_kite OBJ-mesh wing once its tethers are pre-stretched beyond
  ~51 m of standoff): calc_R_t_to_w's tangent-frame azimuthal axis is
  derived from the wing position's tiny horizontal component, whose sign
  can flip from a millimeter-scale change, flipping the whole frame 180°;
  the existing degeneracy guard only caught exact zero. Widened to
  hypot(x, y) < 1e-2 * norm(pos) (within 0.57° of zenith) so the frame is
  deterministic near vertical.
• apply_heading! cannot correct the flip above (or any other upstream
  twist) because its no-ref-points branch re-applied the azimuth/elevation
  rotation to Q_b_to_w, which apply_azimuth_elevation! had already
  rotated — the double-application cancelled the twist out of the measured
  heading, so the solved correction was ≈ 0. Now measures the current
  heading directly from the body's actual world orientation, so the solved
  delta always lands exactly on transform.heading regardless of any twist
  picked up earlier in the transform chain.

Fixed in StableAWEModels 1.0.4, but not here.