Σ_2D = J · Σ_cam · Jᵀ with the general projection Jacobian, as three vec2
columns (TSL has no mat3x2): J[k] = res/2 * (P[k].xy / w - clip.xy * P[k].w
/ w^2), P = cameraProjectionMatrix — GLSL twin gsplatProjectedCovariance.
The Σ_2D entries are materialised UNCONDITIONALLY (they are exactly the
values once first-consumed inside the major-axis branch and read
uninitialised on the near-diagonal path).
Σ_2D = J · Σ_cam · Jᵀ with the general projection Jacobian, as three vec2 columns (TSL has no mat3x2): J[k] = res/2 * (P[k].xy / w - clip.xy * P[k].w / w^2), P = cameraProjectionMatrix — GLSL twin
gsplatProjectedCovariance. The Σ_2D entries are materialised UNCONDITIONALLY (they are exactly the values once first-consumed inside the major-axis branch and read uninitialised on the near-diagonal path).