Why do solutions such as Recast voxelize a 3D scene before transforming it back into polygons?
Would it not be possible to just determine intersections of meshes, check for slope angles and adjust triangulations in that way? I think agent height problems could also be solved by testing sweeps upwards against the other meshes.
DMGregory's point about voxelization's simplicity seems about right. Another reason though, is performance. With these navmesh algorithims, you need a way to reduce the complexity of a scene.
Voxelization is a way of reducing the complexity of high poly meshes. Additionally it has the same effect as spatial partitioning whereby the number of potential neighbors of any feature (triangle/voxel) is reduced to a small, finite amount. You can imagine, depending on the mesh, that a long triangle could require intersection checks with a large portion of the scene, while that will never be the case with a voxel.
If you're interested in different approaches then Recast, here's a video by Casey Muratori in which he explains the (real-time) walking system he developed for the Witness. From what I remember, instead of voxels, he uses sort of poles to quantize the world (thereby simplifying it)
