chore: raise the Python floor to 3.14, and record why it moved back up

[eaitbrahim]

Support only Python 3.14 and above.

This reverses #283, which was a measured decision, so the change is careful to record what shifted rather than just flip a number. The whole argument lives in tests/test_python_floor.py's docstring, which is the file that exists to keep this honest.

The measurement from #283 still stands

3.13, 3.12 and 3.11 all pass identically to 3.14. Nothing in this code needs 3.14. assert_never (3.11+) was the one binding feature, and 3.10 still fails collection on it.

So this is a policy floor, not a measured one, and the docstring says so in those words. The two must not be confused by whoever revisits this.

What changed around the measurement

• Users stopped supplying the interpreter. Decide the Python floor: justify >=3.14.4 or relax it #283 predates the signed desktop bundle (D5) and scripts/install.sh, both of which bootstrap their own environment. The floor now reaches contributors and packagers rather than everyone who installs — a far smaller group, and one that can install a Python.
• 3.14 was days old when Decide the Python floor: justify >=3.14.4 or relax it #283 was decided. Distributions package it now.

What it unblocks — both verified, not assumed

ruff can target py314. It could not before, and ruff.toml recorded exactly why: under py314 against a >=3.11 floor, ruff format rewrote except (A, B, C): into PEP 758's unparenthesised form — a SyntaxError on 3.11–3.13, which the metadata then invited. Raising the floor removes every interpreter that cannot parse it, so the hazard is gone rather than tolerated. ruff check is clean at py314 with no new UP violations.

The numpy.* mypy override is deleted. numpy's bundled stubs use PEP 695 type statements that a python_version = "3.11" run cannot parse, so the workspace silenced the transitive crawl — follow_imports = "skip" plus follow_imports_for_stubs — rather than raise the floor every package pins. mypy now passes clean without it, 359 source files.

CI collapses to one leg

Two legs existed to test the two ends of a span that no longer exists. It stays a matrix rather than a scalar so a second supported version is one list entry, and so the required status context stays test either way.

Three literals became derivations

The installer gate (tests/test_install_script.py), the CI leg, and the docs check all now derive the floor from requires-python instead of repeating it. Raising or lowering it again is one edit plus the docstring.

The docs assertion was tightened because it false-passed. "3.14" in readme matched "any Python 3.11+ (the repo develops on 3.14)" — the development pin satisfied the test while the stated floor stayed wrong. It now requires the 3.14+ form. The Arabic mirror is checked for the version alone, because "3.14 فما فوق" is better Arabic than transliterating the English token.

One test deleted with no successor

test_the_binding_feature_that_sets_the_floor_still_exists asserted assert_never was still imported. That was right for a measured floor of 3.11 and is meaningless at 3.14 — it binds nothing. A policy floor has no binding feature, and asserting one would state a reason that is not the reason. The absence is deliberate and the docstring says so.

Superseded floors stay cited in CONTRIBUTING.md and in the docstrings on purpose, and the docs test deliberately does not scan for them — a scan would punish exactly the record-keeping the contribution guide asks for.

Files

8 × requires-python · ruff.toml · mypy python_version and the numpy override · ci.yml · release.yml · scripts/install.sh · packaging/macos_app.sh · README, README.ar, CONTRIBUTING, docs/desktop-install.md, docs/launch.md · 4 test files · uv.lock

Gates

• ruff check keel tests packages — All checks passed
• mypy — no issues in 359 source files
• pytest -q — 4665 passed, 3 skipped
• uv lock --check — current

Note

If the floor blocks a contributor, that is the argument that would lower it again — and there is no feature standing in the way. CONTRIBUTING.md now says so explicitly and invites the issue.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NyeggYtojNXCTHeD3JHxb6