BUG: Fix uninitialized output in LabelSet erode and dilate at zero scale

[physwkim]

LabelSetErodeImageFilter and LabelSetDilateImageFilter only call AllocateOutputs() and gate every write on m_Scale[d] > 0, so a zero radius on the last axis skips erosion's only output-writing pass and the filter returns uninitialized memory. Addresses B90 of #6575.

Root cause and fix

Erosion commits to the output only on lastpass, defined as the geometrically last dimension. Under UseImageSpacing, a zero radius on that axis makes its scale zero, the axis pass never runs, lastpass is never reached, and the output is whatever AllocateOutputs() left there. With every radius zero, neither filter writes the output at all.

Two changes, both in the shared base: lastpass is now gated on m_LastActiveDimension — the highest-index axis that actually runs a pass — so the commit happens on an axis that exists; and when no axis is active at all, the output is seeded from the input, which is the correct identity result for a zero-radius morphological operation.

The seed is deliberately scoped to the fully-degenerate case. An unconditional seed was measured at 7-10% of GenerateData() wall time on a 182x218x182 label volume and is pure waste, because on any active configuration the lastpass block writes every pixel of the region — verified by poisoning the seed with a sentinel and confirming all six baseline-compared tests still match bit-for-bit.

Local validation

New GoogleTest itkLabelSetMorphBaseImageFilterGTest.cxx in a new driver:

• ErodeZeroRadiusOnLastAxisStillErodesActiveAxis — fail-before: uninitialized garbage. Pass-after: correct erosion on the active axis.
• ErodeAllZeroRadiusIsIdentity / DilateAllZeroRadiusIsIdentity — fail-before: uninitialized garbage (e.g. 147 where 0 was expected). Pass-after: identity.

ctest -L LabelErodeDilate: 8/8 pass, bit-identical baselines.

AI assistance

• Tool: Claude Code
• Role: implementation and regression-test authoring from the analysis in BUG: Findings collected while porting ~60 filters — reported together in one issue #6575
• All code was reviewed, built, and tested locally before committing

[greptile-apps]

Greptile Summary

This PR fixes zero-scale LabelSet morphology output handling. The main changes are:

• Commit erosion output on the last active dimension.
• Copy input to output when no dimension pass runs.
• Add regression tests for partially zero erosion and all-zero erode/dilate cases.
• Add GoogleTest wiring for the new tests.

Confidence Score: 5/5

This looks safe to merge.

No blocking issues found in the changed code.

No files need attention.

T-Rex Logs

What T-Rex did

• I reviewed the general-contract-validation-proof and confirmed it documents command detection, pip install, configure output, partial build output, CTest listing, and an attempt to build the LabelErodeDilate tests.
• I noted that the cmake --build command for LabelErodeDilateGTestDriver and LabelErodeDilateTestDriver did not finish within the allotted budget and was terminated after progressing only through dependency compilation.
• I confirmed that no LabelSetMorphBase zero-radius GTest binary was produced to run tests due to the incomplete build.
• I anchored the proof log as the primary artifact for reviewers to inspect the setup steps and the blocker details.

_{Ran code and verified through T-Rex}

Important Files Changed

Filename	Overview
Modules/Filtering/LabelErodeDilate/include/itkLabelSetErodeImageFilter.hxx	Uses the last active dimension when deciding where erosion writes final output.
Modules/Filtering/LabelErodeDilate/include/itkLabelSetMorphBaseImageFilter.h	Adds state for the last dimension with active scale.
Modules/Filtering/LabelErodeDilate/include/itkLabelSetMorphBaseImageFilter.hxx	Computes active dimensions and seeds identity output for fully inactive morphology.
Modules/Filtering/LabelErodeDilate/itk-module.cmake	Adds the GoogleTest test dependency.
Modules/Filtering/LabelErodeDilate/test/CMakeLists.txt	Registers the new GoogleTest driver source.
Modules/Filtering/LabelErodeDilate/test/itkLabelSetMorphBaseImageFilterGTest.cxx	Adds tests for zero-radius erosion and dilation behavior.

_{Reviews (2): Last reviewed commit: "BUG: Fix uninitialized output in LabelSe..." | Re-trigger Greptile}

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.

[hjmjohnson]

Force-push fixes the m_BaseSigma/NaN dependency left by this PR's own fix. ctest -L LabelErodeDilate: 8/8 pass.

Root cause and fix

firstval (the axis whose scale seeds m_BaseSigma) was keyed on m_Radius[P] != 0, while the per-axis passes and the new m_LastActiveDimension gate are keyed on m_Scale[P] > 0. With UseImageSpacing() at its default of false, the +1 margin means m_Scale[P] is positive even at m_Radius[P] == 0 — so the two conditions diverge. With radius=[0,3]: m_Scale=[1, 5.5], firstval=1 (first nonzero radius) → m_BaseSigma=5.5, but axis 0's pass actually runs with sigma=1. The commit's own new erosion test in itkLabelSetUtils.h is an exact float equality against m_BaseSigma, which the axis-0 pass can never reach — zeroing the entire output. Separately, an all-zero radius with UseImageSpacing on produced m_Scale = [0, NaN] (0.0/0.0), and the new identity-path guard was only correct because NaN > 0 evaluates false — a fragile, undocumented dependency.

Fixed by keying firstval on m_Scale[P] > 0 (the same predicate already used for m_LastActiveDimension), and guarding the normalization divide so it only runs when a nonzero-scale axis was actually found. This makes the identity-path guard (renamed from anyScaleActive, now reusing firstValFound — they were computing the exact same condition) correct by construction instead of by NaN comparison semantics.

Two finder passes in review proposed re-keying on m_Radius[P] != 0 instead of m_Scale; that's disprovable: with UseImageSpacing off and radius=[3,0], m_Scale=[5.5, 1], so axis 1 is active and its pass does run — keying on radius would again pick the wrong axis. m_Scale is the only field that says which axis actually gets a pass.

Also fixed: m_LastActiveDimension had no initializer (indeterminate int on a protected base member until first assigned).

Local validation

New test ErodeWithDefaultSpacingAndZeroRadiusOnFirstAxisIsNotAllZero (default UseImageSpacing=false, radius=[0,3]) — the filter's actual default was previously untested; all three existing tests use UseImageSpacingOn(). Verified fail-before/pass-after by stashing the fix and re-running: fails against the unfixed code (interior pixel reads 0 instead of the label), passes with the fix.

pre-commit run --all-files: exit 0. LabelErodeDilateGTestDriver: 4/4 pass. ctest -L LabelErodeDilate: 8/8 pass, 0 regressions.