Native float ordering diverges from Spark for NaN and signed zero

[ShreyeshArangath]

Describe the bug
Native floating-point comparisons and min/max aggregations diverge from Spark SQL semantics for NaN and signed zero (-0.0 vs 0.0).

To Reproduce
Run the following DataFrame queries:

val maxDouble = Seq(Double.NaN, Double.PositiveInfinity, Double.MaxValue)
  .map(Tuple1.apply).toDF("a").selectExpr("max(a)").first()
assert(java.lang.Double.isNaN(maxDouble.getDouble(0)))

val comparison = Seq((0.0, -0.0)).toDF("pos", "neg")
  .select($"pos" > $"neg").first()
assert(comparison.getBoolean(0) == false)

Actual Auron results:

• Native max(a) returns a non-NaN value (Double.PositiveInfinity) instead of NaN.
• Native 0.0 > -0.0 evaluates to true.

Reproduced on Spark 3.1 and Spark 3.5.

Expected behavior
Spark SQL floating-point ordering semantics:

• NaN is equal to NaN and greater than any non-NaN value.
• -0.0 is equal to 0.0.

max(a) should return NaN when NaN is present, and 0.0 > -0.0 should return false.

Screenshots
N/A

Additional context

• Affected upstream commit: master at 9f90eacb
• Upstream source file and line references:
  • spark-extension/src/main/scala/org/apache/spark/sql/auron/NativeConverters.scala:576-581
  • native-engine/datafusion-ext-plans/src/agg/maxmin.rs:110-124, 207-220
• Distinction from Support NormalizeNaNAndZero #1015 and support normalize nan and zero #1016: those changes added NormalizeNaNAndZero for grouping keys, but direct comparisons and native min/max aggregations still use native float comparisons directly.

[Sigma-Ma]

Hi, I'd like to work on this issue. I plan to reuse the existing NaN/zero normalization for native float comparisons, and make both the partial update and merge paths of min/max follow Spark's float ordering. I'll add Float/Double regression tests for NaN, signed zero, infinity, and NULL, including partition merges.

[ShreyeshArangath]

assigned it to you @Sigma-Ma