CameraPose

Builds a world WGS84-ECEF -> image-UV 4x4 matrix for a camera at a known geographic position looking at a known geographic target. The output matrix takes a world ECEF position to image clip space; downstream code divides xy/w to land in image UV in 0, 1 across the camera's frustum.

Composition (right-to-left): imageProjection = NDC->UV bias * perspective(vFov, aspect) * lookAt(eye, target, up)

The bias term folds the NDC -1, 1 ->0, 1 mapping into the matrix so the fragment stage only needs the homogeneous divide; no additional shader-side rescale.

Always operates in WGS84 ECEF - independent of any 2D display projection (Mercator, polar, ...) the consuming Globe might use for rendering. The vertices the matrix is applied to come from terrain rendering, which keeps real-Earth ECEF positions regardless of display projection.

Construction is allocation-free after the first call: scratch Vec3/Matrix4s and a private Wgs84Projection are held as fields and reused on each setToLookAt invocation.

The view matrix follows the standard right-handed lookAt(eye, target, up) recipe; up is the ellipsoid surface normal at the camera position. Roll isn't applied (negligible for stabilised drone gimbals; the source-image vertical flip lives in the texture's coord transform).