[zfergus]

Description

Moves the Eigen ABI settings (EIGEN_DONT_VECTORIZE and the CUDA alignment pins) from ipc_toolkit and individual dependencies onto the Eigen3::Eigen target. It touches the CMake setup, ipc/utils, the tests and CI.

• cmake/ipc_toolkit/ipc_toolkit_eigen_abi.cmake adds ipc_toolkit_pin_eigen_abi(<eigen_target> [DONT_VECTORIZE] [PIN_ALIGNMENT]). It appends the definitions to the Eigen target's INTERFACE_COMPILE_DEFINITIONS, so every target that links Eigen inherits them.
• The per-dependency defines in finite_diff.cmake and meshfem_sparse.cmake are removed, since those targets now inherit the settings.
• ipc/utils/eigen_abi.hpp adds a runtime check that the caller and the library agree on Eigen's ABI.
• .github/workflows/odr-check.yml links the tests with LTO and -Wodr, and fails on any ODR violation.

Why. Eigen picks its alignment per translation unit from the compile flags. That alignment decides the layout of fixed-size types and which allocator Eigen uses. Before this PR only ipc_toolkit and a hand-patched list of dependencies carried EIGEN_DONT_VECTORIZE.

Built with	EIGEN_MAX_ALIGN_BYTES	sizeof(struct { double; Vector2d; })	sizeof(VectorMax<double, 18>)	Heap allocator
EIGEN_DONT_VECTORIZE (IPC Toolkit)	0	24	152	malloc/free
No flags (a dependency by default)	16	32	160	malloc/free
-march=x86-64-v3, no define	32	32	160	Eigen's offset allocator

• Reproduced in a two-file program (GCC 15.2, Ubuntu 26.04). A mismatched layout gives wrong values and silent buffer overruns in arrays of structs. Returning a fixed-max vector by value fails with *** stack smashing detected ***. A mismatched allocator fails with free(): invalid pointer.
• Why failures look flaky. Whether this crashes depends on which inline Eigen copy the linker keeps and what gets inlined. At -O0 the linker keeps one consistent copy and nothing fails, so passing tests don't rule it out.
• Real instance. On dynamics, polysolve_linear was built without the define. LTO -Wodr found its Vector2d (16-byte aligned) conflicting with the IPC Toolkit's (8-byte aligned) in PointEdgeDistance<double, 2>::point_edge_closest_point_direction.
• Also unpinned on main: tight-inclusion, scalable-ccd and tinyad were built without the define. They don't currently share a mismatched type, but they now inherit the pin.

API changes

• New: ipc::EigenAbi, ipc::eigen_abi(), ipc::library_eigen_abi() and ipc::eigen_abi_matches_library().
  • eigen_abi() and eigen_abi_matches_library() are static on purpose. As inline functions the linker would merge them, and the library would report the caller's values.
• Downstream builds that link ipc::toolkit now get the definitions through Eigen3::Eigen. That also applies to the parent project's other targets that link the same Eigen target.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

How Has This Been Tested?

• [eigen_abi] test passes.
• A consumer compiled without EIGEN_DONT_VECTORIZE and linked against eigen_abi.cpp.o reports a mismatch. With matching flags it reports a match.
• Full test suite: 457/457 on Ubuntu 26.04 (GCC 15.2.0, Release, -march=x86-64-v3). 456/456 on macOS arm64 (Apple clang 21, Debug) before the CMake helper refactor.
• Every third-party target that links Eigen now compiles with EIGEN_DONT_VECTORIZE (finite-diff, MeshFEM, tight-inclusion, scalable-ccd, tinyad). I checked this per target from the build commands.
• The LTO -Wodr build reports 0 violations on this branch. Its check step reports the violation above on a dynamics build without the polysolve fix.
• The new ODR Check workflow has not run on GitHub yet; this PR is its first run.
• CUDA builds: not tested locally. The CUDA alignment pins moved to the Eigen target unchanged.

Test Configuration:

• OS and Version: Ubuntu 26.04.1 (KVM guest, AMD Ryzen 9 5900XT); macOS (arm64)
• Compiler and Version: GCC 15.2.0; Apple clang 21

Checklist

• My code follows the clang-format style guidelines of this project
• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have followed the project style guide
• My changes generate no new warnings
• Any dependent changes have been merged and published in downstream modules

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.86%. Comparing base (182537a) to head (2adb383).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #297   +/-   ##
=======================================
  Coverage   96.86%   96.86%           
=======================================
  Files         227      229    +2     
  Lines       23035    23047   +12     
  Branches     1619     1620    +1     
=======================================
+ Hits        22312    22325   +13     
+ Misses        723      722    -1     

Flag	Coverage Δ
unittests	96.86% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.