[zhouyc-ali]

Purpose

Paimon and Vortex use different NaN comparison semantics. For FLOAT/DOUBLE columns containing NaNs, native filtering can drop matching rows even for finite literals such as value > 0. Residual filtering cannot recover those rows.

Decline floating-point value predicates in the Vortex converter so the caller can evaluate them with Paimon's residual filter. Keep IS NULL / IS NOT NULL and supported AND conjuncts pushed down. Floating-point filters may read more rows as a result.

Tests

• Add FLOAT/DOUBLE regressions for all six comparisons, IN/NOT IN lists of 1/20/21 values, NULL checks, and mixed AND/OR predicates.
• Write/read finite values, signed zeros, infinities, NaNs with different signs/payloads, and NULLs. Verify candidate row IDs retain every expected match and residual-filtered IDs match Paimon filtering of an unfiltered read.
• The four new parameterized test invocations fail without the fix and pass with it.
• Module verify: 73 tests passed; 4 optional Java/Python E2E tests skipped. Checkstyle, Spotless, Maven Enforcer and Apache RAT passed.
• Fix the existing CSV dynamic-option test to trigger validation at scan creation. All 32 CSV/runtime-validation tests pass on JDK 8 and 11.