fix(core): keep NaN semantics of negated float and double comparisons - #3048

❗ Please review the guidelines for contributing

Description

Negated float and double comparisons are decompiled into the opposite comparison, which gives a different result for NaN:

public int notLess(double a, double b) {
    if (!(a < b)) {
        return 1;
    }
    return 2;
}

This is decompiled as if (a >= b) { return 1; }, so notLess(Double.NaN, 1.0) returns 2 instead of 1. The same happens with return !(a > b); (decompiled as return a <= b;) and with ternaries.

cmpg-*/cmpl-* return 1 or -1 when an argument is NaN, so cmp(a, b) op 0 equals a op b only for some ops. Also, once the condition is folded into a op b, the condition inversions that run later (IfOp.invert(), invertIfRegion) are no longer exact.

With this change, a float/double cmp stays wrapped in the condition until code generation, so every inversion still works on the integer cmp result. ConditionGen then writes a op b if that gives the same result for NaN, and !(a inverted_op b) otherwise. Long comparisons are folded as before. Comparisons that were already correct, like the ones in TestCmpOp/TestCmpOp2, decompile to the same code as before.

TestFloatCompareNaN is new. On master, it fails in both the d8-j11 and java-11 profiles (expected: 1 but was: 2 from the auto check). With the fix, ./gradlew test, spotlessCheck and checkstyle pass.

I used Claude Code to write this change.