ENH: Ingest ITKSmoothingRecursiveYvvGaussianFilter into Modules/Filtering

[hjmjohnson]

Ingest the standalone remote module
InsightSoftwareConsortium/ITKSmoothingRecursiveYvvGaussianFilter
into Modules/Filtering/SmoothingRecursiveYvvGaussianFilter/. Vliet/Young/Verbeek
recursive Gaussian smoothing — drop-in alternative to the canonical
itk::SmoothingRecursiveGaussianImageFilter. In-tree under
EXCLUDE_FROM_DEFAULT. GPU classes are gated on ITK_USE_GPU.

Built directly on upstream/main via the v4 ingestion pipeline (#6204).

Commits

#	Subject
1	ENH: Ingest ITKSmoothingRecursiveYvvGaussianFilter into Modules/Filtering (Mode-A merge)
2	DOC: Add SmoothingRecursiveYvvGaussianFilter README and in-tree CMakeLists.txt
3	COMP: Remove SmoothingRecursiveYvvGaussianFilter .remote.cmake (in-tree)
4	ENH: Enable Module_SmoothingRecursiveYvvGaussianFilter in configure-ci
5	STYLE: gersemi-format SmoothingRecursiveYvvGaussianFilter src/CMakeLists.txt

Merge topology preserved: 24 merge commits, 87 commits unique to this PR.

OpenCL handling

The upstream top-level CMakeLists.txt carried bespoke
find_package(OpenCL) + try_run detection that conflicts with ITK's
in-tree GPU support (ITKGPUCommon). It is intentionally not migrated;
a clean three-line project() + itk_module_impl() form takes its
place. GPU sources continue to build under ITK_USE_GPU via
src/CMakeLists.txt.

Phase B follow-up

After merge, archive the upstream repo via the v4 archive script:

Utilities/Maintenance/RemoteModuleIngest/archive-remote-module.sh \
    SmoothingRecursiveYvvGaussianFilter

[hjmjohnson]

@greptile review

[greptile-apps]

Greptile Summary

This PR ingests the ITKSmoothingRecursiveYvvGaussianFilter remote module in-tree as Modules/Filtering/SmoothingRecursiveYvvGaussianFilter/, replacing the remote .cmake descriptor with a standard in-tree layout. GPU sources are gated on ITK_USE_GPU; several issues from the previous review round were resolved in commit 53f020e (raw new[] leaks, undefined INRIA_ITK_LIBRARIES, typo in GPU_DEPENDENCIES, orphaned OpenCLInfo.cpp).

• Core algorithm (CPU path): RecursiveLineYvvGaussianImageFilter and the outer SmoothingRecursiveYvvGaussianImageFilter are clean; SetUp() is called with actual input spacing inside BeforeThreadedGenerateData, giving correct per-dimension Gaussian coefficients.
• GPU path: GPUSmoothingRecursiveYvvGaussianImageFilter::SetSigmaArray reads spacing from this->GetOutput()->GetSpacing() at parameter-set time rather than at execute time, so any SetSigma call before Update() freezes the coefficients at spacing=1.0 regardless of the actual image spacing.
• Build/test: The remote .cmake descriptor is removed and pyproject.toml is updated for CI; add_definitions(-DWITH_DOUBLE) in the non-GPU test path silently compiles tests in double precision while the module itself uses float.

Confidence Score: 4/5

Safe to merge for CPU-only builds; the GPU path computes incorrect Gaussian coefficients for any image with non-unit pixel spacing, which is the normal case for medical images.

The GPU filter computes its recursive filter coefficients from the output DataObject's spacing at SetSigma call time. Before Update() runs, that spacing is the ITK default (1.0), not actual input image spacing. Since GPUGenerateData() never re-invokes SetUp(), the GPU filter silently produces incorrect results for any image with non-unit pixel spacing — the common case for CT, MRI, and microscopy data. The CPU sibling handles this correctly. The module is EXCLUDE_FROM_DEFAULT and GPU is opt-in, limiting exposure, but the defect affects all GPU users.

Modules/Filtering/SmoothingRecursiveYvvGaussianFilter/include/itkGPUSmoothingRecursiveYvvGaussianImageFilter.hxx — the GPU coefficient setup does not use actual input spacing

Important Files Changed

Filename	Overview
Modules/Filtering/SmoothingRecursiveYvvGaussianFilter/include/itkGPUSmoothingRecursiveYvvGaussianImageFilter.hxx	GPU filter computes Gaussian b-coefficients from output spacing at SetSigma time (typically default 1.0), not actual input spacing; produces incorrect kernel for non-unit-spacing images
Modules/Filtering/SmoothingRecursiveYvvGaussianFilter/include/itkGPUSmoothingRecursiveYvvGaussianImageFilter.h	Class declaration with std::vector members replacing raw new[] (fixed in 53f020e); gated behind #ifdef GPU and ITK_USE_GPU correctly
Modules/Filtering/SmoothingRecursiveYvvGaussianFilter/include/itkSmoothingRecursiveYvvGaussianImageFilter.h	CPU outer filter declaration; three typedef declarations should be using aliases to match the rest of the file; otherwise clean
Modules/Filtering/SmoothingRecursiveYvvGaussianFilter/include/itkRecursiveLineYvvGaussianImageFilter.hxx	CPU line filter correctly calls SetUp() with actual input spacing in BeforeThreadedGenerateData; raw new[]/delete[] in ThreadedGenerateData are correctly paired for ProcessAborted but would leak on unexpected exceptions
Modules/Filtering/SmoothingRecursiveYvvGaussianFilter/itk-module.cmake	Module dependencies correctly declared with ITKGPUSmoothing and ITKGPUCommon gated on ITK_USE_GPU; typo GPU_DEPENDANCIES fixed in 53f020e
Modules/Filtering/SmoothingRecursiveYvvGaussianFilter/src/CMakeLists.txt	Correctly compiles only GPU kernel sources under ITK_USE_GPU; undefined-variable target_link_libraries block and unreachable guards dropped in 53f020e
Modules/Filtering/SmoothingRecursiveYvvGaussianFilter/test/CMakeLists.txt	add_definitions(-DWITH_DOUBLE) in non-GPU path means tests compile with double precision while the production code uses float, creating a precision mismatch between tests and real usage
Modules/Remote/SmoothingRecursiveYvvGaussianFilter.remote.cmake	Remote module descriptor correctly removed now that the module is in-tree under Modules/Filtering/
pyproject.toml	Module_SmoothingRecursiveYvvGaussianFilter:BOOL=ON correctly added to configure-ci cmake flags

Sequence Diagram

sequenceDiagram
    participant User
    participant GPUFilter as GPUSmoothingRecursiveYvvGaussianImageFilter
    participant Output as OutputDataObject
    participant GPUKernel as OpenCL Kernel

    User->>GPUFilter: SetSigma(sigma)
    GPUFilter->>Output: GetSpacing() returns [1.0, 1.0] default pre-pipeline
    GPUFilter->>GPUFilter: "SetUp(spacing=1.0) coefficients computed HERE"
    Note over GPUFilter: b0,b1,b2,b3,M-matrix frozen with spacing=1.0

    User->>GPUFilter: "SetInput(imageWithSpacing=0.5mm)"
    User->>GPUFilter: Update()
    GPUFilter->>GPUFilter: GPUGenerateData()
    Note over GPUFilter: BuildKernel() called once no SetUp() call
    GPUFilter->>GPUKernel: "Launch with stale coefficients spacing=1.0"
    Note over GPUKernel: Wrong Gaussian kernel for 0.5mm spacing

    Note right of GPUFilter: CPU path BeforeThreadedGenerateData calls SetUp(actualSpacing)

Loading

_{Reviews (2): Last reviewed commit: "Merge remote-tracking branch 'upstream/m..." | Re-trigger Greptile}

[greptile-apps]

GPU Gaussian coefficients computed with wrong spacing for non-unit-spacing images

SetSigmaArray reads spacing from this->GetOutput()->GetSpacing() (line 195). Before Update() is called, the output DataObject carries default ITK spacing (1.0 in every dimension), not the actual input image spacing. Since GPUGenerateData() never calls SetUp() again, the b-coefficients and M-matrix are frozen at the values computed with spacing=1.0. For any real-world image where pixel spacing != 1.0 mm (e.g., CT with 0.5 mm voxels), the GPU path silently produces an incorrect Gaussian kernel while the CPU sibling (RecursiveLineYvvGaussianImageFilter::BeforeThreadedGenerateData, line 288) correctly calls this->SetUp(pixelSize[m_Direction]) with the actual input spacing every time the filter runs.