[Dilshad-Tanvi]

NumRef's eq and cmp, which back ==, <, sorted, in and dict/set key lookup for every int/float mix, round a big int to f64 before comparing. So (1 << 53) + 1 == float(1 << 53) is True, (1 << 53) + 1 > float(1 << 53) is False, and (1 << 2000) == float("inf") is True because the int rounds to infinity. The spec says mixed int/float comparison is mathematically exact even when neither operand fits the other's type, and starlark-go compares through big.Rat. The float.star conformance lines that check exactly this (p53, the 1.23e45 neighbours, maxint64) were being filtered out in tests/go.rs.

Small ints still compare as f64, which is exact for i32. A big int is compared against floor(f) as a BigInt, with any fractional part breaking the tie, and non-finite floats sort above or below every int. eq now goes through cmp, so equality only gets stricter and equal values still hash the same. The filtered conformance lines now run and pass. #254 left comparisons alone on purpose, and this covers them.

[meta-codesync]

This pull request has been imported. If you are a Meta employee, you can view this in D127353263. (Because this pull request was imported automatically, there will not be any future comments.)