Adopt the py-canon standard, and read Parquet

[soodoku]

Three things that had to happen together, because the standard's linters are what surfaced the first one.

A bug: max_dependence_points did nothing

estimate(max_dependence_points=N) is documented, estimate_from_file forwards it, and _dependence_diagnostics reads it and subsamples above it — but the call inside estimate omitted the argument, so every run used the 2500 default. Asking for a cheaper diagnostic on a large frame set changed nothing.

ruff's ARG001 found it: an accepted, documented, unread parameter. Same shape as the ci_level bug in #12. Two regression tests, both watched failing with the fix reverted:

1 failed, 1 passed   <- fix reverted
2 passed             <- fix restored

CSV is no longer the interface

estimate_from_csv → estimate_from_file, dispatching on the suffix: .parquet/.pq, or .csv/.tsv with optional compression. A new read_frames exposes the reader on its own.

Parquet is what the docs show now, and the reason is not taste: CSV carries no types, so a count column with one missing value comes back as float64 and timestamp comes back as a string something downstream has to guess the format of. Two tests pin this — one asserts the formats give the same estimate to twelve places, the other asserts the dtype difference is real.

No compatibility alias. The package has never been published (PyPI 404s), so nothing imports the old name.

The standard

CI, docs, release and Dependabot now call gojiplus/py-canon reusable workflows instead of carrying copies; python-publish.yml is superseded by release.yml. ruff at the fleet configuration, pyright instead of mypy, pydoclint on docstrings, plus CHANGELOG.md, CITATION.cff, pre-commit and zizmor.

That started at 326 ruff findings. Most were the config landing wrong rather than real problems:

• per-file-ignores pointed at tests/**, but the tests live at geoinference/tests/** — so nothing was exempt. Fixing the path, and adding PT009/PT027 because the suite is unittest.TestCase throughout and rewriting 140 assertions buys nothing, cleared 235.
• allowed-confusables = ["×"]: the codebase writes grid dimensions as 100 × 100, which is what the notation means.
• T201 exempted for io.py and simulate.py, whose printed output is their interface.

The residue was real and is fixed: nine missing docstrings, two missing Raises sections, a lambda-as-def pair, seven over-long lines, and a test that unpacked ratio_ci.wild without checking it wasn't None.

The Python floor rises to 3.12 with the fleet. Never published, so nothing depended on 3.11.

Verification

preen check --strict   21 issues -> exit 0   (one info-level flat-layout note left)
pytest                 76 passed, 1 skipped, 6 subtests
ruff check / format    clean, 19 files
pyright                0 errors, 0 warnings
pydoclint              no violations
twine check            both artifacts PASSED
coverage               84%, CI floor set to 80

Also merged ahead of this: #12 (the ci_level fix, independently re-verified) and #5/#6/#7/#16. Note that #5/#6/#7 bumped actions in ci.yml/docs.yml, which this PR replaces with reusable-workflow callers — so those bumps are now moot, but merging them cost nothing and kept the queue clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YDg9BEDJgFhsixKFjAquNm

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.