Drop the MatIS special case in FunctionSpace.local_to_global_map

[pbrubeck]

Description

FunctionSpace.local_to_global_map() carried a mat_type argument for one reason: a local submatrix of a MATIS reported its local-to-global map in a different layout from one of a MATLOCALREF, so masking the boundary condition nodes had to read a different array for each.

if mat_type == "is":
    indices = lgmap.indices.copy()
    unblocked = False
else:
    # LGMap is already unrolled
    indices = lgmap.block_indices.copy()
    unblocked = True

PETSc !9575 puts MATIS on the same layout as MATLOCALREF, so the scalar indices are in block_indices for both and the branch has nothing left to distinguish. This PR drops it, along with the mat_type argument threaded through ParloopBuilder to supply it.

The MATIS arm was never correct anyway: it read lgmap.indices, which is only the scalar map when the subspace starts on a block boundary of the mixed space. It also never reached the assembled matrix, because ParloopBuilder.replace_lgmaps() skips installing the masked map when m.handle.type == "is". What it did do was keep indices[nodes] = -1 from going out of bounds, which is why it cannot simply be deleted against the currently pinned PETSc.

Testing

test_assemble_matis gains a mixed space whose blocked subspace comes second, so that it starts at a degree of freedom of the mixed space that is not a multiple of its own block size — the case PETSc got wrong. The existing parametrizations put the blocked subspace first, where the offset is zero and the bug does not show.

Against a PETSc that carries !9575, all 10 parametrizations pass on 3 processes. Against the pinned PETSc, 5 of them fail: mixed-is with boundary conditions raises IndexError from indices[nodes] = -1, and both mixed_blocks-is cases raise PETSc error 75 from MatGetLocalSubMatrix().

I also checked, outside the test suite, that every combination of scalar, vector and tensor subspaces in a two-field mixed space assembles identically under mat_type="is" and mat_type="aij", with and without Dirichlet conditions, in serial and on 3 processes. Before !9575, 10 of those 25 combinations either raised or assembled the wrong matrix.

make srclint is clean. It stops at cython-lint, which is not installed in my environment; no .pyx file is touched.

Blocked on PETSc

CI cannot go green until PETSc !9575 reaches a release that SUPPORTED_PETSC_VERSION (currently v3.25.0) points at. !9575 is stacked on !9511, and both need to land in release first.

To build against it before then, the petsc/petsc branch pbrubeck/matis-localref-block-sizes is current release plus the whole stack.

Opened as a draft for that reason.

---

Assisted by Claude Code (Claude Opus 5).

[stefanozampini]

I find this confusing: you have blocked indices (one per block), but the flag is unblocked