Row filter: `<`/`<=` on a column missing from the file keeps all rows instead of matching none

[yuhao-su]

Problem

In the arrow row filter (PredicateConverter, crates/iceberg/src/arrow/reader/predicate_visitor.rs), when a predicate references a column that is not present in the parquet file (schema evolution: the column was added after the file was written), the comparison arms fall back to inconsistent constants:

• less_than / less_than_or_eq → build_always_true() (~L416, ~L436 at 2768ccf8)
• greater_than / greater_than_or_eq / eq / not_eq → build_always_false()

All carry the same comment "A missing column, treating it as null" — but under null semantics every comparison against a missing (all-null) column is non-matching, so the correct constant is always_false for all six.

References

• Spec, Column Projection: a field id not present in the data file reads as null (or its initial-default).
• iceberg-java ParquetMetricsRowGroupFilter#lt (and every other comparison): valueCount == null → "the column is not present and is all nulls" → ROWS_CANNOT_MATCH.

Impact

always_true only over-keeps rows, so engines that re-apply the predicate (RisingWave, DataFusion) still return correct results and merely lose pruning. A consumer relying on the parquet-level row filter being exact gets wrong results for < / <= on evolved columns: every row of every old file passes.

Suggested fix

Change the less_than / less_than_or_eq fallbacks to build_always_false(), and document whether the row filter is exact or advisory for unresolvable references (see the variant handling, where present-but-unresolvable columns are conservatively kept — matching java's ParquetMetricsRowGroupFilter#notNull variant carve-out "Leave these type filters to be evaluated post scan").