GH-4235, GH-4239: NaN in numeric comparison and in the xsd:boolean cast

[anishmehta24]

GitHub issues resolved #4235, #4239

Pull request Description:

Two coupled NaN fixes, one commit each (this PR replaces #4244, which afs noted overlapped with it).

GH-4239 (cf55ae2d) CastXSD.castToBoolean tested for NaN with NodeValue.sameValueAs(nv, nvDoubleNaN), which is never true because NaN is not the same value as anything, so xsd:boolean("NaN"^^xsd:double) fell through to true. The float case only passed by accident ("NaN"^^xsd:float is a different term, so it missed the fast path and hit Double.compare(NaN, NaN) == 0). The numeric branch now returns XSDFuncOp.effectiveBooleanValue(nv), which is exactly the XPath cast rule quoted above the method: zero, -0.0 and NaN are false, every other number true.

GH-4235 (ce7b01c6) compareNumeric uses Double.compare/Float.compare, where NaN is greater than everything, so FILTER(?v > 1e0) accepted NaN. XPath numeric comparison defines every ordering comparison involving NaN as false, = false and != true. In NVCompare:

• compareByValue (behind NodeValue.compare: <, <=, >, >=) returns CMP_INDETERMINATE for a numeric comparison involving NaN, so the operator raises ExprNotComparableException and the filter is an error. The same-term fast path no longer returns CMP_EQUAL for NaN (NaN >= NaN must not be true).
• sameValueAs is false whenever either side is NaN, so "NaN"^^xsd:double = "NaN"^^xsd:float is false and != true (already the case for two NaN terms of the same datatype via sameExceptNaN, which now shares an isNaN helper).
• compareWithOrdering (compareAlways, ORDER BY) keeps NaN's place after every other number exactly as before, rather than falling through to term ordering.

The two are coupled: making sameValueAs strict about NaN is what exposes the accidental float-NaN cast result, so the cast fix has to land with or before the comparison fix.

Tests: TestCastXSD.cast_to_boolean_06a/06b/06c; TestExpressions2.nan_cmp_01..12; TestNodeValue.testCompareNaN_value / testCompareNaN_ordering. Nine of them fail on main. The full jena-arq module passes (16629 tests; the single error is TestLangJsonLD_prefixes fetching a remote context, unrelated).

---

• Tests are included.
• Documentation change and updates are provided for the Apache Jena website (not needed)
• Commits have been squashed to remove intermediate development commit messages (one commit per issue).
• Key commit messages start with the issue number (GH-xxxx)

By submitting this pull request, I acknowledge that I am making a contribution to the Apache Software Foundation under the terms and conditions of the Contributor's Agreement.

🤖 Generated with Claude Code