COMP: Expose itkeigen subdir in ITKEigen3_INCLUDE_DIRS for external consumers

[hjmjohnson]

Add ${ITKEigen3_SOURCE_DIR}/src/itkeigen to ITKEigen3_INCLUDE_DIRS so external consumers using the upstream Eigen include convention (<Eigen/<header>>) resolve against the vendored Eigen via ITK's IMPORTED targets — currently they get fatal error: Eigen/Dense: No such file or directory. Unblocks InsightSoftwareConsortium/ITKTotalVariation # 57's modern-target-interfaces refactor for proxTV. Independent of and complementary to #6176 (the Eigen 5 update); the architecture issue predates that PR.

Root cause

ITK ships its vendored Eigen3 headers at:

Modules/ThirdParty/Eigen3/src/itkeigen/Eigen/<header>

ITK's own internal code uses ITK_EIGEN(<header>) from itk_eigen.h, which expands to <itkeigen/Eigen/<header>>. With include path ${ITKEigen3_SOURCE_DIR}/src (the parent of itkeigen/), the compiler appends /itkeigen/Eigen/<header> and resolves correctly. ITK proper has always built fine.

Modules/ThirdParty/Eigen3/CMakeLists.txt:58 (vendored case, before this PR):

set(ITKEigen3_INCLUDE_DIRS ${ITKEigen3_SOURCE_DIR}/src)

That single path flowed into INTERFACE_INCLUDE_DIRECTORIES of both ITK::ITKEigen3Module and ITK::eigen_internal. External consumers using the upstream <Eigen/<header>> convention got the wrong root and failed to resolve. The Eigen3::Eigen IMPORTED target produced by Eigen's own export does have the correct path, but it's declared as add_library(Eigen3::Eigen ALIAS eigen) — an in-source-scope alias, invisible to FetchContent sub-builds (verified empirically).

Fix

Single-block change to Modules/ThirdParty/Eigen3/CMakeLists.txt:58:

-  set(ITKEigen3_INCLUDE_DIRS ${ITKEigen3_SOURCE_DIR}/src)
+  set(
+    ITKEigen3_INCLUDE_DIRS
+    ${ITKEigen3_SOURCE_DIR}/src
+    ${ITKEigen3_SOURCE_DIR}/src/itkeigen
+  )

ITK's own <itkeigen/Eigen/<header>> consumers continue to find their headers via the first entry. External consumers using <Eigen/<header>> find theirs via the second. Both flow into INTERFACE_INCLUDE_DIRECTORIES of the IMPORTED ITK::ITKEigen3Module so the fix applies in-tree and to find_package(ITK) consumers identically.

Local verification

Full ITK build with Module_TotalVariation=ON pinned at ITKTotalVariation # 57's tip (39e849e, which previously failed lapackFunctionsWrap.cpp:11:10: fatal error: Eigen/Dense: No such file or directory):

Step	Result
Configure (ITK_BUILD_DEFAULT_MODULES=ON, Module_TotalVariation pinned at PR # 57 tip)	✅ clean
Build ninja -j16	✅ 6165/6165 targets, 0 failures
libitkproxTV-6.0.so produced	✅
bin/TotalVariationTestDriver linked	✅
ctest -R ProxTV	✅ itkProxTVImageFilterTest passes
lapackFunctionsWrap.cpp.o compile flags include -isystem .../Eigen3/src/itkeigen	✅ (the vendored path resolves <Eigen/Dense>)
Compile flags do not include -isystem /usr/include/eigen3	✅ (no longer falling back to system Eigen)

Compile-commands receipt for lapackFunctionsWrap.cpp:

-I/tmp/itk-eigen-include-fix-build/_deps/proxtv_fetch-src/src
-I.../Eigen3/src
-I.../build/Modules/ThirdParty/Eigen3/src
-isystem .../Eigen3/src/itkeigen        ← the new include dir from this PR

Relationship to other PRs

• ENH: Update vendored Eigen3 to 5.0.1+master; verify USE_SYSTEM 3.3/3.4/5.x #6176 (Eigen 5 vendored update) — independent. The architecture issue this PR fixes predates the Eigen 5 update; upstream/main has the same setup. PR ENH: Update vendored Eigen3 to 5.0.1+master; verify USE_SYSTEM 3.3/3.4/5.x #6176 is good to merge as-is. After both this PR and ENH: Update vendored Eigen3 to 5.0.1+master; verify USE_SYSTEM 3.3/3.4/5.x #6176 land, no further coordination is needed.
• InsightSoftwareConsortium/ITKTotalVariation# 57 (cmake_interface_update) — unblocked by this PR. Their set(proxTV_Eigen_LIBRARIES ITK::ITKEigen3Module) becomes correct as-is, no further changes needed in TotalVariation. The longstanding "silent fall-through to system Eigen" bug in ITKTotalVariation main (where set(Eigen3_DIR "${ITKInternalEigen3_DIR}") reads an empty value) becomes moot once # 57 lands, because proxTV no longer needs the broken Eigen3_DIR shim.

[hjmjohnson]

Tracking issue for the SYSTEM-include macro cleanup is now open: #6224 — captures the macro-plumbing diagnosis, the three approaches in increasing scope, and the recommended sequencing (investigate the existing eigen_internal SYSTEM INTERFACE first; only touch ITKModuleMacros.cmake if needed).

[blowekamp]

This comment is too long, and not efficient for humans to understand.

[blowekamp]

PR #57 was not blocked because of a compilation error it was blocked because of incompatible ITK versioning infrastructure.

[seanm]

I haven't bisected yet, but looking at what changed in ITK since yesterday, I suspect this change broke my build:

In file included from /some/path/ITK-install/include/ITK/itkAffineTransform.h:21:
In file included from /some/path/ITK-install/include/ITK/itkMatrixOffsetTransformBase.h:24:
In file included from /some/path/ITK-install/include/ITK/itkTransform.h:24:
In file included from /some/path/ITK-install/include/ITK/itkSymmetricSecondRankTensor.h:29:
/some/path/ITK-install/include/ITK/itkSymmetricEigenAnalysis.h:23:10: fatal error: 'itkeigen/Eigen/Eigenvalues' file not found
#include ITK_EIGEN(Eigenvalues)
         ^~~~~~~~~~~~~~~~~~~~~~

[seanm]

Bisect results:

There are only 'skip'ped commits left to test.
The first bad commit could be any of:
bc87d16
64ddc66
We cannot bisect more!

[seanm]

So in fact it was MR #6176 I'll move comments there...