Compute GPU FPFH descriptors from resident radius neighborhoods

.github/workflows/ci.yml

@@ -182,6 +182,10 @@ jobs:
             args: -p spatialrust-registration --features register-fpfh
           - name: spatialrust-registration-fpfh-meta
             args: -p spatialrust --features register-fpfh
+          - name: spatialrust-registration-fpfh-gpu
+            args: -p spatialrust-registration --features register-fpfh-gpu
+          - name: spatialrust-registration-fpfh-gpu-meta
+            args: -p spatialrust --features register-fpfh-gpu
           - name: spatialrust-metrics
             args: -p spatialrust-metrics --features metrics-distance
           - name: spatialrust-metrics-meta

CHANGELOG.md

@@ -21,6 +21,13 @@ removed no sooner than the next major (see `docs/API_STABILITY.md`).
 
 ### Changed
 
+- Add opt-in `register-fpfh-gpu` descriptor computation: reusable radius
+  neighborhoods feed GPU SPFH and distance-weighted FPFH passes, with explicit
+  sample/tree/query upload and descriptor/status readback receipts. Input and
+  neighborhood ownership are checked; incomplete/malformed rows invalidate
+  the entire device result and return errors at download. CPU FPFH/RANSAC
+  defaults remain unchanged; GPU angular/bin rounding is documented.
+
 - Add explicitly uploaded reusable GPU query batches and resident KD-tree
   nearest-k/radius neighborhoods under `gpu-wgpu`. k=0..=64 retains CPU
   selection and stable distance order; radius rows retain traversal order and

README.md

@@ -653,8 +653,16 @@ GPU k-NN/radius search can retain its results for later device passes with
 `GpuQueryBatch::upload` and
 `GpuKdTree::{nearest_k_resident,radius_neighbors_resident}` (`gpu-wgpu`). Radius
 capacity overflow is counted explicitly and complete readback rejects it.
-This provides the neighborhood foundation for feature compute; FPFH/ISS/MLS
-remain CPU algorithms. See [contracts and measurements](notes/2026-10-12_gpu_bounded_neighborhoods.md).
+This provides the neighborhood foundation for feature compute. See
+[contracts and measurements](notes/2026-10-12_gpu_bounded_neighborhoods.md).
+
+FPFH descriptors have an explicit GPU path under `register-fpfh-gpu`:
+`fpfh_descriptors_gpu(&runtime, &cloud, radius, neighbor_capacity)` runs radius
+search, SPFH and weighted FPFH on device and reports all payload transfers.
+Resident APIs reuse uploaded input and neighborhoods. Capacity overflow returns
+an error; GPU rounding can change angular bins near boundaries. CPU matching/
+RANSAC and ordinary descriptor defaults are retained; ISS/MLS remain CPU.
+See [FPFH measurements](notes/2026-10-12_gpu_fpfh_descriptors.md).
 
 ### Library