[harrism]

Adds resource forms of the point-set overloads of tools::cuda::pointsToGrid and voxelsToGrid, completing the set started in #2374. Like the single-set resource forms, they take (sets, stream, resource). Each set's grid, the merged result and all internal scratch allocate through resource on stream, and the result is a GridHandle<cuda::Buffer<std::byte, ResourceT>>, defaulting to DeviceResource:

std::vector<std::tuple<const Coord* const, size_t, double>> sets{{d_a, nA, 1.0}, {d_b, nB, 0.5}};
auto handle = nanovdb::tools::cuda::voxelsToGrid<float, const Coord*>(sets, stream);       // DeviceResource
auto viaRef = nanovdb::tools::cuda::voxelsToGrid<float, const Coord*>(sets, stream, ref);  // borrowed resource

The point-set prototype forms are constrained the same way as the single-set ones, so a cudaStream_t argument selects the resource form. So does a stream wrapper that converts to cudaStream_t, such as at::cuda::CUDAStream; before, the wrapper was deduced as the prototype buffer and failed deep inside the build, in the single-set forms too. A literal 0 in that position now fails with a message saying to write cudaStream_t(0).

Empty inputs give empty outputs, as in MeshToGrid, the topology tools and merge/split, so a pipeline doesn't have to read a device count back before building:

• Builders: a zero point count builds a valid empty grid in every form, including the PointsToGrid builder and DistributedPointsToGrid. An empty list of sets gives an empty handle, and a set without points contributes an empty grid. Before, a zero count launched a zero-block kernel, which exits the process in debug builds and builds on garbage in release builds.
• Checksum: updateChecksum with CheckMode::Full handles an empty grid the same way the host does.
• Merge and copy: mergeGridHandles of no grids and copyTo of an empty handle return an empty handle over the prototype's resource for single-space buffers, instead of throwing for a resource that can't be default-constructed.
• Validator: tools::checkGrid no longer asserts in debug builds on a grid with no nodes. That assert also fired for host-built empty grids.

Invalid arguments still throw std::invalid_argument: a negative count, a bad voxel size or map, and now PointType::Disable for a point grid, which used to throw only after the grid was allocated.

This also fixes pointsToGrid with PointType::Default for pointer-to-const coordinates (e.g. const Vec3f*). It failed to compile in every form, because it assigned into a const element type; the other element-type uses in the file already strip the const.

This targets master. Whether it goes into 13.2 is still open: it's additive and nothing in 14.0 depends on it.

Validation on one RTX PRO 6000 Blackwell GPU with nvcc 12.6:

• New test: TestBuffer.ResourceFormsOfThePointSetOverloads covers the voxel and point set forms with a counting resource, the default resource, byte equality with the prototype forms, the per-set point counts, balanced allocations, empty and mixed set lists, and the invalid-voxel-size and PointType::Disable errors. Its point sets use const pointers, so it also covers the PointType::Default fix.
• Empty inputs: ZeroPointsBuildAnEmptyGrid, EmptyHandlesMergeAndCopyForEveryBufferType and StreamWrapperSelectsTheResourceForm (all in TestBuffer), plus ZeroPoints_DistributedCudaPointsToGrid run single-GPU. The empty grids pass isValid(CheckMode::Full), and a -G debug build of the device and host paths also validates them.
• Existing suites: all single-GPU ctest targets pass. Unit tests, tools, examples and the Python module build with -Werror, and the 38 Python tests pass.

🤖 Generated with Claude Code

[swahtz]

Thanks, this fills in the last of the point-set overloads nicely, and the const PtrT fix for PointType::Default is a good catch. I have three things I'd like addressed before merge. A set with zero points still gets through the new empty check. A stream wrapper type can select the prototype form. And one of the new throws has no test. I also left two smaller suggestions, on the duplicated empty check and on taking the set vector by const&.

[swahtz]

Looks good, thanks for addressing my notes.