Support Python 3.11–3.14 and NumPy 2

[pscicluna]

Support Python 3.11–3.14 and NumPy 2

I'm not entirely sure if Starfish is still alive or if it is now abandonware, but I find it useful and I'm not really aware of a good alternative, so I thought I would take a shot at updating some of the key infrastructure for it. This is primarily intended to fix issue #160, and in doing so it seemed logical to make a few other updates to modernise the package. The updates were written by Claude, but are fairly simple - I've reviewed it all and can't see any obvious problems, and would welcome a thorough review. Claude's own summary of the changes is below.

Problem

Starfish currently cannot be installed on any Python newer than 3.10:

• python_requires = ">=3.7,<3.11" makes pip reject the package outright on 3.11+
  (ERROR: Package 'astrostarfish' requires a different Python: 3.11.5 not in '<3.11,>=3.7').
• On older Pythons, the capped dependencies (astropy>=4,<6, numpy<2, …) force pip to
  build very old sdists from source, which fails against modern build tooling (e.g. the
  astropy 4.x ImportError: cannot import name 'soft_unicode' from 'markupsafe' reported
  in Installation problems through pip #160).
• Two dependencies are effectively dead: nptyping==1.* (abandoned; incompatible with
  NumPy 2 and the source of the <3.11 cap) and the dataclasses PyPI backport (only
  ever needed on Python 3.6, and harmful on modern interpreters).

Since Starfish is pure Python, unblocking modern interpreters is almost entirely a
packaging/metadata fix plus a handful of small code changes.

Changes

Packaging

• Migrated all metadata from setup.py to a PEP 621 pyproject.toml;
  requires-python = ">=3.11", classifiers for 3.11–3.14.
• Removed the dataclasses backport and nptyping.
• Relaxed dependency caps to floors: astropy>=5.3, numpy>=1.24 (NumPy 1.x and
  2.x are both supported), scipy>=1.10, scikit-learn>=1.2, extinction>=0.4.7
  (ships wheels through cp314), h5py>=3.8, flatdict>=4.0, toml>=0.10.2,
  tqdm>=4.60.
• toml is kept (rather than stdlib tomllib) because SpectrumModel.save() relies on
  toml.TomlNumpyEncoder for writing NumPy scalars, and tomllib is read-only.

Code compatibility

• Type annotations moved from nptyping to numpy.typing.NDArray
  (spectrum.py, emulator/emulator.py, models/utils.py) — annotation-only.
• np.trapz → scipy.integrate.trapezoid (7 sites; same algorithm, works on
  NumPy 1.24–2.x) and np.Inf → np.inf (removed in NumPy 2).
• eval("extinction.{law}") → getattr(extinction, law) (law names were already
  whitelist-validated).
• plt.style.use("seaborn") → "seaborn-v0_8" with a graceful fallback (the bare name
  was removed in matplotlib 3.6, so importing emulator.plotting raised).
• Raw-stringed two matplotlib labels that trigger SyntaxWarning: invalid escape sequence '\l' on Python 3.12+ (an error in a future Python).
• Removed the dead "samplers" entry from Starfish.__all__ (no such module).
• Fixed test_no_par_truncation, which called np.asarray() on the PHOENIX grid's
  ragged parameter list — an error since NumPy 1.24. The test now iterates the list
  directly (equivalent to the old object-array behaviour).

CI / docs

• CI matrix bumped to Python 3.11–3.14 × {ubuntu, macos, windows}; updated stale
  actions (checkout@v4, setup-python@v5, cache@v4, codecov@v4); dropped the
  unmaintained pytest-black plugin from the test run.
• Release workflow now uses python -m build and
  pypa/gh-action-pypi-publish@release/v1 (replaces the deprecated pep517.build).
• Docs: app.add_stylesheet → app.add_css_file (removed in Sphinx 4), unpinned the
  docs toolchain (builds cleanly on Sphinx 9.1), migrated .readthedocs.yml to the
  current build.os/build.tools schema.

Correctness and performance

No numerical code path changes behaviour: the annotation swap has zero runtime effect,
scipy.integrate.trapezoid is the implementation NumPy's own migration notes point to,
and np.Inf/np.inf are the same object.

Verified explicitly by running the old code + old pinned deps (Python 3.10 / NumPy 1.26)
against the new code (Python 3.13 / NumPy 2.5) on identical inputs across 13 outputs
(all transforms, all five extinction laws, PHOENIX flux normalization): the
trapezoid-affected paths are bit-for-bit identical; the remainder agree to
≤ 7×10⁻¹³ relative (FFT differences between NumPy builds, not this diff). The
pytest-benchmark suite shows performance parity.

Testing

• Full test suite passes on all 12 CI jobs — Python 3.11 / 3.12 / 3.13 / 3.14 ×
  ubuntu / macos / windows — against real downloaded PHOENIX models:
  https://github.com/pscicluna/Starfish/actions/runs/29474810935
• Additionally run locally on Python 3.11 against both NumPy 1.26 and NumPy 2.4
  with identical results, confirming the change is NumPy-1/2 agnostic.
• python -m build + twine check pass; clean pip install verified from both wheel
  and sdist on 3.11–3.14.

🤖 Generated with Claude Code