[bueler]

Description

ImplicitMatrixContext.createSubMatrix rebuilds each DirichletBC with bc.reconstruct(field=row_inds, ...). For a space that is not a subspace of a mixed space, fs.index is None, so BCBase.as_subspace returns None for field=[0], and the BC is silently dropped. The submatrix then has ordinary operator rows on the boundary, where the original matrix has identity rows.

This happens whenever a general (non-stride) IS covers the whole space; PETSc's identity-stride shortcut in MatCreateSubMatrix avoids it. For example, SNESVINEWTONRSLS with mat_type: matfree and an empty active set does this: CG on the reduced system stagnates or diverges. That is how it showed up, in a FAS multigrid solver for an unconstrained problem.

The fix treats a space with no index as field 0 of itself. BCs on whole mixed spaces are already rejected, so this does not change any mixed-space behaviour. A regression test compares matrix-free and AIJ submatrices over a whole-space general IS, for scalar, vector and vector-component BCs.

AI assistance: diagnosis, fix and test were developed with Claude Code (Claude Opus 5.5) and reviewed by the author.

[bueler]

Could a maintainer please add the "LLM used" label? I apparently don't have permission to set labels.