Fix Python 3.13 test failures (RXMC NumPy 2 bug; document/skip NequIP)

[k-yoshimi]

Overview

Make the test suite pass on Python 3.13. The Test for abICS workflow had been
auto-disabled by GitHub for inactivity; after re-enabling it, two pre-existing
Python 3.13 failures surfaced. Both are unrelated to the recent MLIP work; they
affect develop directly.

1. RXMC replica exchange crash under NumPy 2 (real bug — fixed)

abics/sampling/rxmc.py::find_procrank_from_Trank returned the raw
np.argwhere(...) result — a 2-D array of shape (1, 1) — which was then passed
as the MPI source/dest rank to comm.Recv/comm.Send. NumPy 2 rejects int()
on a non-0-dimensional array, raising

TypeError: only 0-dimensional arrays can be converted to Python scalars

so any replica exchange crashed on Python 3.13 (numpy>=2). NumPy 1.x silently
allowed the size-1 conversion, which is why Python 3.9 passed.

Fix: return a plain int(i[0, 0]), and fix the dead if i is None guard
(np.argwhere never returns None) to if i.size == 0.

Reproduced and verified locally with Python 3.13 / numpy 2.4.6 / mpi4py 4.1.2:
the mock, potts, and potts_pamc sampling tests and the 31 unit tests pass.

2. NequIP cannot run on Python 3.13 (upstream limitation — documented & skipped)

The supported nequip series (<0.7, pinned in pyproject.toml) calls the
removed numpy.in1d, so it needs numpy<2. Python 3.13 has no numpy 1.x wheels,
so this path cannot run there. This is an upstream nequip/numpy-2 incompatibility,
not an abICS bug.

• Exclude ActiveLearnNequip from the Python 3.13 CI matrix (nequip is still
  tested on Python 3.9 with all numpy variants).
• Pin numpy<2 in install_nequip.sh (before and after the nequip stack).
• Document the numpy<2 / no-Python-3.13 requirement for both the NequIP and
  Allegro install paths (JA/EN), and explain the nequip <0.7 cap in
  pyproject.toml, plus developer notes in the workflow and install script.

Also

• Re-enable the Test for abICS workflow and add a workflow_dispatch trigger
  (it had been auto-disabled by GitHub).

Follow-up

Restoring numpy-2 / Python 3.13 coverage for NequIP requires migrating the abICS
nequip solver/trainer to nequip >=0.7 (a backwards-incompatible rewrite),
tracked separately.

Review

Reviewed with Codex over three rounds; converged with no remaining blockers.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR updates abICS to pass CI on Python 3.13 by fixing a NumPy 2 incompatibility in RXMC replica exchange, and by documenting/skipping the NequIP integration path where upstream constraints prevent Python 3.13 support.

Changes:

• Fix TemperatureRX_MPI.find_procrank_from_Trank to return a plain int rank (avoids NumPy 2 scalar-conversion failure during MPI send/recv).
• Document the NequIP/Allegro numpy<2 requirement (and Python 3.13 incompatibility) and enforce numpy<2 in the NequIP install script.
• Update the GitHub Actions matrix to skip the NequIP integration test on Python 3.13 and re-enable manual workflow runs.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
abics/sampling/rxmc.py	Fixes MPI rank handling under NumPy 2 by returning a scalar int from np.argwhere results.
tests/integration/active_learn_nequip/install_nequip.sh	Pins numpy<2 before/after NequIP stack installation to avoid upstream NumPy 2 incompatibility.
pyproject.toml	Adds developer-facing rationale for capping nequip <0.7 and its NumPy/Python constraints.
docs/sphinx/en/source/tutorial/other_models.rst	Documents numpy<2 and Python 3.13 limitation for NequIP/Allegro install paths.
docs/sphinx/ja/source/tutorial/other_models.rst	Same as above in Japanese docs, including numpy<2 guidance.
.github/workflows/Test_abICS.yml	Skips NequIP integration on Python 3.13 and adds workflow_dispatch trigger.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.