If you lay out a cubemap as a 2d texture that looks literally like https://www.turais.de/content/images/size/w1000/2021/05/Stan... it's not hard, given the VU1-based triangle processing (like proto-mesh-shaders 25 years ago), to set the UVs of triangles independently to use the correct square even in the case of dynamic reflections. This doesn't do per-pixel spherical UV normalization. But, with a dense enough mesh, a linear approximation looks good enough.
Except... The triangle UVs will often cross over between multiple squares. With the above texture, it will cross over into the white area and make the white visible on the mesh. So, you fill the white area with a duplicate of the texels from the square that is adjacent on the cube. That won't work for huge triangles that span more than 1.5 squares. But, it's good enough given an appropriate mesh.
If you lay out a cubemap as a 2d texture that looks literally like https://www.turais.de/content/images/size/w1000/2021/05/Stan... it's not hard, given the VU1-based triangle processing (like proto-mesh-shaders 25 years ago), to set the UVs of triangles independently to use the correct square even in the case of dynamic reflections. This doesn't do per-pixel spherical UV normalization. But, with a dense enough mesh, a linear approximation looks good enough.
Except... The triangle UVs will often cross over between multiple squares. With the above texture, it will cross over into the white area and make the white visible on the mesh. So, you fill the white area with a duplicate of the texels from the square that is adjacent on the cube. That won't work for huge triangles that span more than 1.5 squares. But, it's good enough given an appropriate mesh.
Probably would have been better to just use a lat-long projection texture like https://www.turais.de/content/images/size/w1600/2021/05/spru... Or, maybe store the cubemap as independent squares and subdivide any triangles that cross square boundaries.