[crvernon]

Stacked PR — part of the Demeter v3.0.0 upgrade. Base: v3/out-1.3-transitions. Review and merge in order, bottom of the stack first.

Plan items: 4.1, 4.3, 4.5, 4.6, 1.9b, 5.2 (partial)

Benchmark: IDENTICAL on the pinned stack. Latest stack (NumPy 2 / pandas 3) is WITHIN_TOLERANCE (4.7e-10 km²). 28/28 tests pass on the minimum, pinned, and latest stacks. This is the first PR that runs the new CI workflow on GitHub. Removing tracked artifacts (build/, comp_data/, pickles) is left as a follow-up.

Commits

Package with pyproject.toml, test old to new dependency stacks, modernize CI

Packaging (plan 4.1, 4.4, 4.5, 4.6):

• PEP 621 pyproject.toml; setup.py is now a shim.
• Version 3.0.0.dev0. Requires Python >= 3.11 (3.10 reaches end of
  life in October 2026).
• gcamreader and matplotlib move to the [gcam] and [plot] extras and
  are imported lazily, with an install hint. Also new: [test], [docs],
  and [dev] extras.
• Comparison pickles and duplicate .pkl files are excluded from
  distributions. A clean build is a 7.4 MB wheel.
• MANIFEST.in no longer references a missing directory and now includes
  the mapping inputs it left out.
• Docker uses python:3.12-slim with a JRE for gcamreader; the old image
  ran Python 3.6, below the minimum setup.py required.

Dependency compatibility (plan 1.9):

• pandas 3 Copy-on-Write made the region array read-only, which broke
  the Taiwan remap. It is now copied.
• inplace=True is replaced throughout (deprecated in pandas 3).
• The projected land test input is added as land_df.csv.gz, an exact
  round trip of the pickle. The pickle files are kept for now (their
  removal is a separate decision). benchmarks/cases.read_projected
  prefers the CSV.
• Test suite (28 tests) passes on three stacks: minimum (numpy 1.26.4,
  pandas 2.1.4, scipy 1.11.4, xarray 2024.1.1); pinned (1.26.4 / 2.2.1 /
  1.12.0); and latest (numpy 2.4.6, pandas 3.1.0, scipy 1.17.1,
  xarray 2026.9.0, with no warnings).
• Benchmark on the latest stack: WITHIN_TOLERANCE of v3.0.0-b4, max
  4.7e-10 km2, about 15% faster. On the pinned stack: IDENTICAL.

CI (plan 4.3, 0.5):

• Runs on pull requests.
• Lint (syntax errors and undefined names).
• Fast tests on ubuntu and macOS, Python 3.11-3.13.
• Full tests on the minimum and latest stacks.
• Benchmark job comparing the global and Texas cases against the
  baseline named in benchmarks/baselines/CURRENT. The compare reports
  go to the job summary and an artifact. The job fails on DRIFT unless
  the PR promotes a new baseline.
• compare.py falls back to state hashes and region x metric totals when
  a baseline has no full arrays.

🤖 Generated with Claude Code