Validate custom hierarchical-FMG / SBR suite under numpy 2 (needs custom petsc4py 3.25 rebuilt vs numpy 2)

[lmoresi]

Context

numpy 2 support landed in #301 (#247). CI is green, but it runs on the conda-petsc env (petsc4py 3.24.5) and does not exercise the custom hierarchical-FMG / triangle-splitting suite (test_1015/1016/1017, level_2/tier_b).

Gap

The custom-MG code targets petsc4py 3.25 (the custom amr build). It uses PETSc.KSP.DMActive.OPERATOR (custom_mg.py:433), a 3.25 enum absent from conda petsc4py 3.24.5:

AttributeError: type object 'petsc4py.PETSc.KSP' has no attribute 'DMActive'

So the suite cannot be validated in the conda dev env — it needs petsc4py 3.25. The only env with 3.25 is the custom amr build, whose petsc4py is currently compiled against numpy 1.26 and is ABI-broken under numpy 2.5 (numpy.dtype size changed).

The suite passes under amr-dev on numpy 1.26 (petsc4py 3.25) — this is purely a numpy-2 verification gap, not a known break.

To close

Rebuild the custom PETSc 3.25 / petsc4py against numpy 2 (petsc-local-build), then run test_1015/1016/1017 in amr-dev under numpy 2 (planned on the cluster, where the custom PETSc + numpy 2 coexist).

Optional hardening (separate)

If the custom-MG path should also run on conda petsc4py, guard custom_mg.py:433 with a fallback: PETSc.KSP.DMActive.OPERATOR (3.25) → ksp.setDMActive(False) (older signature).

[lmoresi]

Resolved via #304 (raises the conda-petsc floor to >=3.25). Root cause was purely the petsc4py version (conda 3.24.x lacked the PETSc.KSP.DMActive API the custom FMG uses), not numpy 2.

With the conda dev env re-solved to petsc4py 3.25.3 + numpy 2.5.0, the full custom hierarchical-FMG / SBR / rotated-free-slip suite passes: test_1015/1016/1017/1018/1020 — 29 passed. So the custom AMR is confirmed numpy-2 clean, validated locally on the conda env (no custom-PETSc rebuild needed for this). Will close when #304 merges.

[lmoresi]

Closed by #304 (merged): conda-petsc floor raised to >=3.25 (+ a sqlite>=3.46 floor to keep the linux-64 solve sane). The conda dev env now runs petsc4py 3.25.3 / numpy 2.5, and the full custom hierarchical-FMG / SBR / rotated-free-slip suite (test_1015/1016/1017/1018/1020, 29 passed) is green — custom AMR confirmed numpy-2 clean, validated on the conda env (no custom-PETSc rebuild required).