[gdevenyi]

Makes TimeVaryingBSplineVelocityFieldImageRegistrationMethod give the same result with DataObject::GlobalReleaseDataFlag on as off. Stacked on #7095; only the last commit (8bf17af) is new.

Root cause

With the flag on, every internal filter releases its inputs after it runs. Several of those inputs are read again later. Turning the flag on in a small 2-D registration and fixing the first failure each time found:

1. TimeVaryingBSplineVelocityFieldTransformParametersAdaptor feeds the transform's control point lattice to one VectorIndexSelectionCastImageFilter per component. The first selector releases the lattice, and the second throws "region is outside of buffered region".
2. TimeVaryingBSplineVelocityFieldTransform::IntegrateVelocityField feeds the lattice to BSplineControlPointImageFilter, which releases the transform's parameters.
3. The registration's fixed and moving resamplers take the virtual domain as a reference image, and they take the level images as inputs. Both get released. The metric then throws "Virtual domain and moving transform displacement field must have the same size".
4. The forward and inverse integrators share the sampled velocity field. The forward integrator releases it, and the inverse integrator then integrates an empty field. This throws nothing: the inverse displacement field just comes out all zero.
5. The fixed mask resampler releases the mask image that the metric still uses. The run then segfaults in ImageMaskSpatialObject::IsInsideInObjectSpace.

Fix

Each of those filters now gets a graft, which shares the buffer, so no data is copied and results are unchanged. The registration uses #7095's MakeGraft helper. Its resamplers now set their geometry with SetOutputParametersFromImage(virtualDomainImage), the same change #7095 made in SyN.

Sites left as they are, because nothing reads them again after the release:

• the inverse integrator's input, since the sampled field is a local that is discarded after the inverse integrator runs;
• the AddImageFilter input in UpdateTransformParameters, since the old lattice is replaced by the sum on the next line. The registration later reads only its metadata.

The multi-metric resampler branch got the same change as the single-metric branch. The tests run only the single-metric path, as in #7095.

Tests

There are two new cases in #7095's itkSyNImageRegistrationMethodReleaseDataGTest.cxx: a two-level 32x32 time-varying B-spline registration, unmasked and masked. Each compares the flag-on displacement and inverse fields exactly (bitwise) with a flag-off run, and checks that the result is nonzero.

• With only BUG: Make SyN and velocity field registration work with GlobalReleaseDataFlag #7095 (this commit's three headers reverted), both new cases fail with the lattice exception (site 1).
• With this commit, all 7 cases in the driver pass. ctest for the ITKRegistrationMethodsv4, ITKDisplacementField and ITKRegistrationCommon labels: 65/65 passed.
• Mutations. Each graft was removed in turn, and the new cases failed every time:

Removed	Result
adaptor selector graft	exception: region outside buffered region
B-spline filter lattice graft	exception: region outside buffered region
forward integrator graft	inverse field differs (0 vs 0.00115)
resampler level-image grafts	fields differ from the first value
SetOutputParametersFromImage (back to the reference image)	exception: virtual domain size mismatch
mask resampler graft	segfault in the masked case

AI assistance

Tool: Claude Code (claude-opus-5-5). Role: found the failing sites by running with the flag on, wrote the fix and tests, and ran the mutation checks. All code was reviewed and tested locally before committing.

[gdevenyi]

@ntustison could you take a look when you have a moment? This makes TimeVaryingBSplineVelocityFieldImageRegistrationMethod (and its transform and parameters adaptor) give identical results with GlobalReleaseDataFlag on, by grafting the images that internal filters would otherwise release while they are still in use.

[gdevenyi]

The Pixi-Cxx (macos-15) failure is unrelated: itkMCI_DSC_case_2_binary in the MorphologicalContourInterpolation remote module segfaulted; that test does not use the time-varying B-spline code this PR changes. Could a maintainer rerun that job?

[gdevenyi]

After the rebase onto the new #7095 (c968cc5), the Azure ITK.Windows and ITK.macOS builds show as failed, but Azure reports both as canceled, with no logs. The GitHub Actions jobs that have finished pass. Could a maintainer rerun the two Azure builds?

[gdevenyi]

The ARMBUILD-x86_64-rosetta failure is unrelated: ImageToVideoFilterTest could not read its baseline HeadMRVolume_Frame0.mha ("Could not create IO object for reading file"), so the test data was missing on the runner. The same job passes on #7095, which this PR is stacked on. Could a maintainer rerun it?

[gdevenyi]

@hjmjohnson, could you rerun the failed ARMBUILD-x86_64-rosetta job (https://github.com/InsightSoftwareConsortium/ITK/actions/runs/41748468814/job/126307315016)? I can't rerun jobs myself. ImageToVideoFilterTest could not find its baseline HeadMRVolume_Frame0.mha on the runner, and the same job passes on #7095. Every other check passes.