Fix unit check

[xylar]

Compare units by what they mean, not how they are spelled

Fixes #8

The problem

The units check compared the units attribute with the data request as plain strings, so a file carrying kg m^-2 s^-1 was flagged with an error telling the modeler it "should be kg m-2 s-1" — two spellings of the same unit. CF requires only that the attribute be "a string that can be recognized by the UDUNITS package", and UDUNITS recognises both, so this was a false positive. It is not a hypothetical one: PISM writes m^2 precisely because pint does not recognise m2. The data request is itself of two minds about this — it writes m^2 for iareagr and iareafl and m-2/s-1 everywhere else — so the checker was demanding the caret in one place and rejecting it in another.

The fix

_parse_units reduces a units string to a canonical (scale, factors) form, where factors is a sorted tuple of (base unit, exponent) pairs. Comparing those canonical forms makes the check insensitive to everything UDUNITS treats as the same unit: the spelling of exponents (m2, m^2, m**2), the separator between factors (space, ., *, middle dot), / for negative exponents, and the order of the factors. Division follows UDUNITS' left-to-right associativity, so kg/m2 s is kg m-2 s rather than kg m-2 s-1.

The parser is deliberately small, and says so. Parentheses, timestamp offsets such as days since 1850-01-01, and decimal scale factors all return None, and _units_match then falls back to the literal string comparison that was there before. The change can therefore only ever accept more units than it did, never fewer, and nothing the parser does not understand is guessed at. Unit prefixes are not something it knows about: km still fails against m, as it should.

A file spelled exactly as the request still logs the unchanged The unit is correct: <units> line — which is what leaves the golden reference log untouched — while an equivalent spelling logs The unit is correct: m2 (equivalent to the requested m^2), so the difference is recorded rather than passing silently.

Testing

23 new tests, 31 in total, all green. Five run the whole checker over a file whose units attribute has been rewritten to an equivalent spelling (m2, m**2, kg s^-1, kg/s, kg.s-1) and require zero errors; one requires that m^3 where m^2 was asked for is still an error, with its message unchanged. The rest table-test _units_match directly, covering the equivalences above, the cases that must not match (m3 vs m2, km vs m, M vs m, kg/m2 s vs kg m-2 s-1), and the strings the parser declines to interpret.

The golden-log regression test passes unchanged, confirming that no existing, exactly-spelled file changes its log output.

Run locally against the latest environment (python 3.14, xarray 2026.7); CI covers the floor environment as well.

Also updated

The check summary at the top of isschecker/checker.py and the corresponding bullet in README.md, so the documented behavior matches what the checker now does.