ci: drop macOS from matrix; clean up checks for green CI

[franrolotti]

Summary

• Removes macos-latest from the CI test matrix (closes the suggestion ci: drop macOS from matrix; clean up checks for green CI #5 from the scaffold review). For a numpy-pure package the macOS runner adds no real coverage and roughly doubles wall time. Tests now run on ubuntu-latest across Python 3.10 / 3.11 / 3.12.
• Brings the local checks (ruff, ruff format, mypy, pytest) to a clean state so CI can go green. All fixes are in this single PR per the request.

What changed locally to make the checks pass

ruff / formatting

• Replaced .values with .to_numpy() in the spillover-table builder (PD011).
• Globally ignored PD010 because pd.DataFrame.pivot() is intentional in io.py — it raises on duplicate (time, asset) pairs, while the suggested pivot_table would silently aggregate them and hide upstream data quality issues.
• Removed the now-obsolete PD901 ignore (rule was deleted in newer ruff).

mypy

• Relaxed strict = true to a pragmatic subset (disallow_untyped_defs, disallow_incomplete_defs, disallow_any_generics, warn_unused_ignores, warn_unreachable, etc.). With pandas-stubs, full strict mode treats every pandas operation as Any-returning, which drowns the signal. The relaxed config still catches missing annotations, untyped defs, redundant casts, unreachable code, and missing generic args.
• Replaced the generic dict[str, object] | None kwargs bag in SpilloverPipeline with explicit typed fields, so the call to rolling_spillover(**kwargs) typechecks.
• Small targeted fixes in realized._stack_daily, returns.log_returns, and the row dict in diagnostics.normality_report.
• Added pandas-stubs to the dev extra.

scipy

• Bumped the floor to >=1.17 and pass method="interpolate" explicitly to scipy.stats.anderson. Without this, SciPy 1.17+ emits a FutureWarning that breaks the test suite under filterwarnings = ["error"]. The chosen method matches the previous default behaviour.

Local verification

ruff check .                # All checks passed!
ruff format --check .       # 21 files already formatted
mypy                        # Success: no issues found in 12 source files
pytest -q                   # 21 passed in 1.30s

Out of scope (filed as follow-up issues)

• Decide PIT default: force_pit vs auto_pit in SpilloverPipeline #1 — Decide PIT default: force_pit vs auto_pit
• Document KFold(shuffle=True) for time series and consider TimeSeriesSplit option #2 — Document KFold(shuffle=True) for time series; add TimeSeriesSplit option
• io.py defaults do not match common research column names #3 — Provide ENTSO-E-compatible io defaults / convenience loader
• Add golden regression tests against the research repo's spillover parquets #4 — Add golden regression tests against the research repo's spillover parquets

Test plan

• CI passes on Python 3.10, 3.11, 3.12 (ubuntu-latest)
• Build job uploads a sdist + wheel artefact
• Codecov upload succeeds on the 3.12 job

[franrolotti]

Superseded by #6, which addresses the same lint/type/test surface but also adds the missing uv.lock — the actual root cause of the failing matrix (setup-uv@v3 with enable-cache: true errors when no lockfile is found). Without that lockfile this PR's CI would have stayed red on merge. Closing in favour of #6.