Holistic review round 2: fix all findings, release 1.7.1

[wlaur]

Patch release fixing everything surfaced by the second holistic pre-release review (10 findings: 1 major, 3 minor, 6 nits — all reproduced at runtime before fixing). One commit per finding.

Fixes

Major

• Raise the typeguard floor to 4.1.3 — typeguard 4.1.0–4.1.2 cannot be imported on Python 3.14 (they use ast.Str, removed in 3.14), so the declared floor >=4.1 was uninstallable on a supported Python. Bisected empirically: 4.1.3 is the first working release.
• Run the dependency-floors job on 3.13 and 3.14 — the floors job ran on 3.13 only, which is exactly why the typeguard floor slipped through: a pure-Python wheel installs on any Python, so --only-binary cannot catch a floor that resolves everywhere but only works on some interpreters. The 3.14 leg was pre-verified locally: resolves typeguard 4.1.3 / pydantic 2.12.0 / pydantic-settings 2.0.0, and the installed suite passes (770 passed, 10 skipped).

Minor

• Derive the non-strict ordering operators from the strict ones — on polars containers, nan >= nan returned True (polars' total order calls NaN equal to NaN) while == said False, breaking the documented ge ⟺ gt ∨ eq consistency. >=/<= are now derived as strict ∨ eq from the same tolerant __eq__. Re-fuzzed ~172k value pairs per operator across all four magnitude containers; (nan, nan) added to the cross-container test matrix.
• Reject Series-vs-Expr magnitude comparisons like the other mixed containers — this pair bypassed the mixed-container guard and fell through to exact (non-tolerant) comparison. Now raises with an Expr-appropriate hint (evaluate the Expr side, or lift the other side with pl.lit).
• Reject a NaN density when convert_volume_mass gets a Polars input — a NaN rho sentinel would flow into the Series/Expr result as NaN, violating the library-wide "missing = null, never NaN" convention for polars results. Now rejected with a message pointing at the null spelling.

Nits

• Point the __hash__ tolerance note at the (rtol, atol) predicate — the comment referenced np.isclose, which is not the predicate in use.
• Attribute the arithmetic inference to the operator overloads in the README — the overview bullet claimed Quantity.__new__ overloads cover * / ** inference; they are separate arithmetic-operator overloads.
• State the actual versioning policy instead of claiming semver (README + usage docs).
• List water in the coolprop README file overview.
• Refresh the polars-expr-coolprop notebook from a clean top-to-bottom run — sequential execution counts, stale stdout removed, profiler display output (which embedded machine paths) dropped.
• Document the public module constants on the API reference — attribute docstrings for all nine encomp.coolprop constants plus EAGER_PLUGIN_MIN_SIZE, CONSTANTS, SETTINGS, CUSTOM_DIMENSIONS. This surfaced a latent docs bug: a comment line starting with # type: in units.py was parsed by sphinx's ModuleAnalyzer as a PEP 484 type comment, silently dropping the attribute docs for the whole module. UNIT_REGISTRY is deliberately left undocumented — autodoc's signature formatter crashes on its UnitRegistry[Any] annotation under -W.

Release chores

• Bump to 1.7.1 (pyproject + uv.lock).
• Serve the README logo from main instead of the release tag, so the version lives only in pyproject.toml.

Gates (after the final commit)

• pyright: 0 errors · pyrefly: 0 errors (46 suppressed) · ruff check + format: clean
• pytest: 957 passed (was 948; new tests cover the three minor fixes)
• sphinx -T -W --keep-going (RTD parity): build succeeded, constants render on the API page

🤖 Generated with Claude Code