[Vedin]

Stacked on #138 (base fix/cleanup-failed-writes); retarget to main once #138 merges.

Manifest bounds were turned into column statistics and pruning arrays as if they were exact extremes of the column. Four cases broke that assumption, and two of them give wrong results.

1. String and binary bounds are truncated. Under the default truncate(16) metrics mode, the lower bound is a prefix of the smallest value and the upper bound an incremented prefix. As exact statistics, DataFusion answered MIN/MAX with the truncated prefixes. They are now Inexact.

2. Float upper bounds exclude NaN (Iceberg spec), while Snowflake orders NaN above every number. Over {1.0, NaN, 1.0} Snowflake gives COUNT_IF(k > 2) = 1, MAX = NaN and order 1, 1, NaN. Trusting the upper bound broke all three:

   • k > 2 pruned the file;
   • MAX came from statistics as 1.0;
   • min = max turned the column into the constant 1.0.

   A float column's upper bound is now used only when the file's nan_value_counts entry is recorded as 0, in both the statistics and planning-time data-file pruning. A float partition's upper bound is used only when contains_nan is Some(false). Lower bounds stay usable.

3. Value::into_any boxed the OrderedFloat wrapper. The bound arrays downcast to f32/f64, so every float bound was null and float columns never pruned files or manifests at planning time. It now boxes the plain float, which enables that pruning. Item 2 keeps it NaN-safe.

4. Zero float bounds are not exact extremes. Parquet writers record a zero minimum as -0.0 and a zero maximum as +0.0 whatever the sign of the stored zeros. MIN over {0.0, 1.0} came back as -0 (Snowflake: 0). Such bounds are now Inexact.

Since float bounds now reach the planning-time predicates, manifest and data-file pruning keeps every container when the predicate fails to evaluate on the bounds (as Parquet row-group pruning does), instead of failing the query.

Cost:

• Files written by this fork record no NaN counts, so their float columns lose upper-bound pruning and MAX from statistics.
• String MIN/MAX now scan.

Not changed: a NaN with the sign bit set (-'NaN', inf - inf) sorts below every number in arrow's total order, unlike Snowflake. That is a separate, pre-existing execution difference.

Tests:

• New unit tests for the bound statistics, both pruning paths and into_any. The NaN pruning tests fail with the guards removed.
• cargo test -p datafusion_iceberg -p iceberg-rust passes, except integration_trino_rest, which fails starting its Trino container in this environment.
• The Rustice end-to-end tests that pin the Snowflake results land with the pin bump.

🤖 Generated with Claude Code