[finsberg]

Moves io4dolfinx's DOLFINx cross-version workarounds to fenicsx-compat, as was done for scifem. It also fixes the 10 tests that fail on DOLFINx nightly on main.

Important

Depends on fenicsx-compat>=0.3.0.

Migration

• io4dolfinx.compat removed. The following now come from fenicsx_compat:
  • cmap, dofmap
  • cell permutation info
  • index map creation and unwrapping
  • index_to_dest_ranks, which always returns (ranks, offsets), so the AdjacencyList/tuple branching in writers.py is gone.
• Inline duplicates replaced:
  • mesh creation with a ghost mode in read_mesh
  • the C++ finite element constructor
  • reconstruct_mesh (still deprecated)
  • the PyVista snake_case setting (4 places)
  • the ADIOS2 scope
• Dead pre-0.10 branches removed, each checked against the DOLFINx v0.10.0 sources:
  • the block_size element constructor
  • the value_shape FunctionSpace constructor
  • create_mesh without max_facet_to_cell_links
  • element.signature() as a method
  • the petsc_options_prefix probe in the legacy reader tests
• Closure dofs per entity are now len(layout.entity_closure_dofs(dim, 0)). That is the value num_entity_closure_dofs gave on 0.10/0.11, and it exists on every version. DOLFINx main removed num_entity_closure_dofs.
• Dependencies: fenicsx-compat>=0.3.0 is added and packaging moves to the test group, since only the tests still use it.
• README: the claim of v0.9.0 support (already ruled out by the >=0.10.0 floor) is gone, and the README now points to fenicsx-compat.

Nightly fixes

Failing tests on main	Cause	Fix
6 × test_vtkhdf.py: "VTK hexahedra need at least 8 nodes"	FEniCS/dolfinx#4621: get_vtk_cell_type takes the node count, not the dimension	fenicsx_compat.vtk_cell_type(cell_type, num_dofs_per_cell)
2 × test_manifold_hcurl_roundtrips_but_transfer_raises[*-RT]	FEniCS/dolfinx#4625: RT on a manifold needs facet orientations	the test calls fenicsx_compat.create_facet_orientations before creating the spaces
2 × test_submesh_roundtrip_codim0_vector[*-2-RT]: error 9.4e-7 vs 1e-13	The test reference field contained sin, which isn't in RT2. On nightly, interpolating it into RT2 gives dofs that differ by about 1e-6 between mesh numberings.	The reference vector field is now linear, so it lies in RT2

The last one is a test bug, not an I/O bug. Outside the suite, the RT2 function read back from file has the same norm (to 15 digits) and the same L2 error against the exact field as the one written. With a field that lies in the space, the re-interpolated DOFs match to 1e-16.

Testing

Run locally against DOLFINx 0.12.0.dev0:

	main	this branch
serial	2466 passed, 10 failed	2476 passed, 24 skipped
mpirun -n 2	2090 passed, 10 failed	2100 passed, 400 skipped
mpirun -n 3	—	2100 passed, 400 skipped

• pre-commit run --all-files and mypy (in the DOLFINx environment, as CI runs it) are clean.
• Not run locally: the stable leg, and the legacy-data tests, which need data generated with legacy DOLFIN.

🤖 Generated with Claude Code