Update vertex geometry in transform code

crates/fj-core/src/operations/transform/cycle.rs

@@ -1,4 +1,5 @@
 use fj_math::Transform;
+use itertools::Itertools;
 
 use crate::{
     operations::insert::Insert,
@@ -20,15 +21,83 @@ impl TransformObject for (&Handle<Cycle>, &Handle<Surface>) {
     ) -> Self::Transformed {
         let (cycle, surface) = self;
 
-        let half_edges = cycle
+        let half_edges_and_old_vertex_geometries = cycle
             .half_edges()
             .pairs()
-            .map(|(half_edge, _)| {
-                (half_edge, surface)
-                    .transform_with_cache(transform, core, cache)
+            .map(|(half_edge, next_half_edge)| {
+                let vertex_a_geom = core
+                    .layers
+                    .geometry
+                    .of_vertex(half_edge.start_vertex())
+                    .unwrap()
+                    .local_on(half_edge.curve())
+                    .unwrap()
+                    .clone();
+                let vertex_b_geom = core
+                    .layers
+                    .geometry
+                    .of_vertex(next_half_edge.start_vertex())
+                    .unwrap()
+                    .local_on(half_edge.curve())
+                    .unwrap()
+                    .clone();
+
+                let half_edge = (half_edge, surface)
+                    .transform_with_cache(transform, core, cache);
+
+                (half_edge, vertex_a_geom, vertex_b_geom)
             })
             .collect::<Vec<_>>();
 
+        // That we're transforming the vertex geometry here, instead of down in
+        // the vertex transform implementation, presents an inconsistency in the
+        // transform architecture. But if we were to transform it down there,
+        // we'd have to pass all kids of information down there (the half-edge
+        // transform code would need the half-edge's end vertex, and we would
+        // need to pass the curve into the vertex transform code).
+        //
+        // Doing it here is easier. It might be a hack, but so be it. I have two
+        // reasons for accepting this, instead of doing it "the right way" (if
+        // what I've described in the previous paragraph actually is the right
+        // thing to do):
+        //
+        // - At some point, it will no longer be necessary to define vertex
+        //   geometry redundantly in every local curve space. Then this whole
+        //   situation will be easier to handle.
+        // - With geometry living in a separate layer now, it's unclear if the
+        //   current approach to transforms is the right one anyway. So far,
+        //   we're still creating new topological objects with each
+        //   transformation, but whether this is desirable is not obvious.
+        //
+        // When we have progress on those two points, whatever the "right way"
+        // is will have changed anyway, so I'm not too worried about a hack here
+        // or there.
+        let half_edges = half_edges_and_old_vertex_geometries
+            .into_iter()
+            .circular_tuple_windows()
+            .map(
+                |(
+                    (half_edge, vertex_a_geom, vertex_b_geom),
+                    (next_half_edge, _, _),
+                )| {
+                    // We have only transformed the vertices in 3D space. They
+                    // still have the same positions in local curve space. We
+                    // just have to copy those over to the new vertices.
+                    core.layers.geometry.define_vertex(
+                        half_edge.start_vertex().clone(),
+                        half_edge.curve().clone(),
+                        vertex_a_geom,
+                    );
+                    core.layers.geometry.define_vertex(
+                        next_half_edge.start_vertex().clone(),
+                        half_edge.curve().clone(),
+                        vertex_b_geom,
+                    );
+
+                    half_edge
+                },
+            );
+
         Cycle::new(half_edges).insert(core)
     }
 }