[vortex] Do not push unrepresentable timestamp predicate literals

[jackylee-ch]

Purpose

toTimestampLiteral takes the precision from the column type but the value from the predicate literal, and nothing reconciles them — Spark builds the literal from its own type with Timestamp.fromMicros, so a TIMESTAMP(0) column can receive a millisecond literal. Dividing by 1000 (or discarding the sub-millisecond part on TIMESTAMP(3)) then yields a bound that answers some operators wrongly: f_ts < 1500ms on a seconds column pushed < 1s and dropped the 1000ms row, which no post-scan filter can restore.

Math.floorDiv is not enough. For a residual literal, col < L needs the bound rounded down but col > L needs it rounded up, and = is unsatisfiable while != is a tautology — one direction cannot serve all six operators. So a literal that is not exactly representable at the column's grain is no longer pushed, and the existing null handling drops the leaf. ParquetFilters.normalizeDecimal uses RoundingMode.UNNECESSARY for the same reason.

Tests

Three roundTrip cases in VortexPredicateConverterTest: millisecond literal on TIMESTAMP(0) with < and !=, a pre-epoch literal, and a sub-millisecond literal on TIMESTAMP(3). All three return 0 rows on master.

Written with Claude Code; verification is mine.

[JingsongLi]

This prevents real false-negative pruning: a lower-precision Vortex timestamp column can receive a higher-precision Spark literal, and truncating that literal changes strict comparisons. I traced null conversion through the leaf and compound visitors: unsupported AND leaves are omitted conservatively, while an OR with an unsupported child is not pushed. That preserves matching rows for residual evaluation.

No blocking issue found in the changed guards, including pre-epoch values and sub-millisecond/microsecond remainders. The added real-format round-trip tests exercise the lost-row cases. I have not rerun the native Vortex suite here.