Split extxyz into ASE-free core + ase-extxyz plugin (v0.3.0)

[jameskermode]

Breaking change for v0.3.0 — drops the ASE dependency from extxyz and adds a sibling ase-extxyz plugin package that wires the C parser into ase.io via the external I/O plugin entry point.

What

Package	Lives	Depends on	Public API
extxyz	python/extxyz/	numpy, pyleri	Frame, iread_dicts, read_dicts, write_dicts, cextxyz (ctypes)
ase-extxyz	python/ase-extxyz/	ase>=3.23, extxyz>=0.3.0	None for direct use; registers cextxyz format with ase.io

After this lands users do:

# C parser only (no ASE)
import extxyz
for frame in extxyz.iread_dicts('traj.xyz'):
    ...

# Via ASE
import ase.io
atoms = ase.io.read('traj.xyz', format='cextxyz')
ase.io.write('out.xyz', atoms, format='cextxyz')

The pre-0.3 top-level names (extxyz.read, extxyz.write, extxyz.iread, extxyz.ExtXYZTrajectoryWriter) raise ImportError with a message pointing at ase-extxyz. ExtXYZTrajectoryWriter is kept on the new package, now keeping a single libc FILE* open across .write() calls (the v0.2.x version held a Python file handle and silently fell back to pure-Python — strictly slower than this).

How

• python/extxyz/extxyz.py (~870 lines, mixed) split into grammar.py (parser AST + Properties + encoders) and core.py (Frame + iread_dicts/write_dicts).
• python/extxyz/utils.py deleted; SinglePointCalculator glue + Atoms ↔ Frame translation moved to python/ase-extxyz/ase_extxyz/io.py.
• python/extxyz/cli.py rewritten to emit JSON via ExtXYZEncoder so the extxyz console script works in a no-ASE install.
• All ASE-using tests (tests/test_kv_parsing_*, test_ase_cases, test_traj_writer) git mv-d to python/ase-extxyz/tests/ and updated to import from ase_extxyz.io / conftest.read. The original tests/conftest.py followed.
• New tests/test_core.py covers the dict API plus a subprocess check that import extxyz does not pull in any ase.* module.

Plugin contract details

Format name: cextxyz (can't shadow ASE's built-in extxyz/xyz).
IOFormat code: +S — multi-frame, string filename. The C parser opens via libc; if ASE pre-opened a Python handle (+F) the C path would be unreachable.
ext intentionally not set — users opt in with format='cextxyz', no auto-detection clash with the built-in.

CI

• New .github/workflows/ase-extxyz.yml: matrix tests + sdist/wheel publish on ase-extxyz-v* tags.
• python-package.yml updated to install both packages and run both pytest suites (with USE_FORTRAN=T for the fextxyz round-trip).
• build-wheels.yml unchanged; its tags: ['v*'] glob doesn't match ase-extxyz-v*, so the two release pipelines stay independent.

Verified locally (macOS, Python 3.13, ASE 3.28)

• ASE-free venv: 10/10 tests/ pass, pip list shows no ase after pip install ./.
• Both-installed venv: 48 pass, 2 skipped (pre-existing 2D string-array cases). assert 'cextxyz' in ase.io.formats.ioformats succeeds. Round-trip via ase.io.write/read matches positions to 1e-7.

Release sequencing (after merge — not part of this PR)

1. Tag v0.3.0 on master → build-wheels.yml publishes extxyz==0.3.0 to PyPI.
2. Wait ~1 min for PyPI to index it.
3. Tag ase-extxyz-v0.1.0 on the same SHA → ase-extxyz.yml publishes ase-extxyz==0.1.0, depending on extxyz>=0.3.0.

Both tags can be cut from the same commit; they just need to land in order so pip install ase-extxyz can resolve the extxyz dep.

Test plan

• CI Python 3.10–3.13 build green on python-package.yml (both extxyz + ase-extxyz tests, plus Fortran round-trip)
• CI Python 3.10–3.13 build green on ase-extxyz.yml
• build-wheels.yml still green for the extxyz package (16 wheels)

🤖 Generated with Claude Code