Besides the comments above, this looks good to me. I was a bit to lazy to get into that -.-
Good to know that tree.visitActiveBBox exists.

This seems reasonable, however have you tried using Tree::voxelizeActiveTiles? Since we have a MaskGrid it will not do any memory allocation, just updating the node's active masks. Then we won't have to have a special case for tiles in the code. I just tried it and it works on the bug report file, maybe there is a special case that I am missing.

diff --git a/intern/cycles/render/volume.cpp b/intern/cycles/render/volume.cpp
index 64196cdd8a8..ac96fd1fb34 100644
--- a/intern/cycles/render/volume.cpp
+++ b/intern/cycles/render/volume.cpp
@@ -285,6 +285,8 @@ void VolumeMeshBuilder::create_mesh(vector<float3> &vertices,
   vector<int3> vertices_is;
   vector<QuadData> quads;
 
+  topology_grid->tree().voxelizeActiveTiles();
+
   generate_vertices_and_quads(vertices_is, quads);
 
   convert_object_space(vertices_is, vertices, face_overlap_avoidance);

Well, that's a lot simpler. If that fixes the issue and passes the tests then you can go ahead and commit that to 2.91. I guess it's a bit more memory in that it still creates the actual leaf nodes, but it should be negligible.

In D9557#237655, @Brecht Van Lommel (brecht) wrote:

Well, that's a lot simpler. If that fixes the issue and passes the tests then you can go ahead and commit that to 2.91. I guess it's a bit more memory in that it still creates the actual leaf nodes, but it should be negligible.

Indeed, I was thinking about the actual voxel buffers. Allocating a bunch of empty leaves is still peanuts compared to that.

Anyway, the tests are passing, and AFAICT there is no side effects, so I am gonna commit this little fix.