…::cast_to

Add comprehensive documentation explaining that struct casting uses field name
matching rather than positional matching. This clarifies the behavior change
for struct types while preserving existing documentation for other types.

Addresses PR review recommendation #4 about documenting public API changes.

Replace .clone() with Arc::clone() to address clippy warning about
clone_on_ref_ptr. This makes the ref-counting operation explicit and
follows Rust best practices.

Fixes clippy error from rust_lint.sh.

## Problem
The PR to fix struct casting (issue apache#14396) introduced regressions where struct
casting with field additions/removals was failing, and name-based field matching
wasn't working correctly in all scenarios.

## Root Causes Identified
1. **Field index mismatch**: cast_struct_array_by_name was using field indices from
   the DataType instead of the actual StructArray, causing wrong column access when
   field names didn't match physical layout.

2. **Missing fallback logic**: When source and target had no overlapping field names
   (e.g. {c0, c1} → {a, b}), name-based matching failed silently, returning NULLs.
   Added fallback to positional casting for non-overlapping fields.

3. **Optimizer const-folding issue**: ScalarValue::cast_to_with_options was calling
   Arrow's cast_with_options directly, which doesn't support struct field count
   changes. The optimizer's simplify_expressions rule would fail when trying to
   fold struct casts at compile time.

4. **Validation rejection**: The logical planner's can_cast_types check rejected
   struct-to-struct casts with mismatched field counts before execution. Added
   special handling to allow all struct-to-struct casts (validation at runtime).

## Solution
- Created datafusion/common/src/struct_cast.rs with shared name-based struct
  casting logic for both runtime (ColumnarValue) and optimization-time (ScalarValue)
- Updated ScalarValue::cast_to_with_options to use the name-based struct casting
- Updated ColumnarValue::cast_to to use the shared logic
- Updated Expr::cast_to validation to allow struct-to-struct casts
- Added fallback to positional casting when field names don't overlap
- Fixed struct array field access to use actual StructArray fields, not DataType fields
- Updated tests to reflect new behavior and correct syntax issues

## Behavior Changes
Struct casts now work correctly with:
- Field reordering: {b: 3, a: 4} → STRUCT(a INT, b INT) → {a: 4, b: 3}
- Field additions: {a: 1} → STRUCT(a INT, b INT) → {a: 1, b: NULL}
- Field removals: {a: 1, b: 2, c: 3} → STRUCT(a INT, b INT) → {a: 1, b: 2}
- Fallback to positional casting when no field names overlap

## Files Modified
- datafusion/common/src/lib.rs: Added struct_cast module
- datafusion/common/src/struct_cast.rs: New shared struct casting logic
- datafusion/common/src/scalar/mod.rs: Use name-based struct casting
- datafusion/expr-common/src/columnar_value.rs: Delegate to shared casting logic
- datafusion/expr/src/expr_schema.rs: Allow struct-to-struct casts through validation
- datafusion/sqllogictest/test_files/struct.slt: Fixed tests and added new ones

…ype comparison

This aligns the type coercion logic with the new name-based struct casting
semantics introduced for explicit CAST operations in issue apache#14396.

Changes:
- struct_coercion in binary.rs now attempts name-based field matching first
- Falls back to positional matching when no field names overlap
- Requires matching field counts for successful coercion
- Preserves left-side field names and order when using name-based matching

This fixes cases where CASE expressions with structs having the same fields
in different orders now correctly match fields by name rather than position.

Note: Several CASE expression tests with struct field reordering are disabled
due to const-folding optimizer hang when evaluating struct literals with
different field orders. This requires separate investigation and fix.

…-folding

This fixes the TODO tests in struct.slt that were causing optimizer hangs
when attempting to const-fold struct literal casts with field count mismatches.

Changes:
1. Modified expr_simplifier.rs can_evaluate() to skip const-folding for
   struct CAST/TryCAST expressions where source and target have different
   field counts. This prevents the optimizer from attempting to evaluate
   these at plan time, deferring to execution time instead.

2. Modified cast.rs cast_with_options() to allow all struct-to-struct casts
   at physical planning time, even when Arrow's can_cast_types rejects them.
   These casts are handled by name-based casting at execution time via
   ColumnarValue::cast_to.

3. Uncommented and fixed TODO tests in struct.slt:
   - CAST({a: 1} AS STRUCT(a INT, b INT)) - adds NULL for missing field b
   - CAST({a: 1, b: 2, extra: 3} AS STRUCT(a INT, b INT)) - ignores extra field

The fix ensures that:
- Optimizer doesn't hang trying to const-fold unsupported struct casts
- Physical planner accepts struct-to-struct casts with field count changes
- Execution uses name-based casting to handle field reordering and NULLs

This uncomments and fixes the TODO tests for struct coercion in CASE expressions
that were previously disabled due to concerns about optimizer hangs.

Changes:
1. Uncommented the TODO test section for struct coercion with different field orders.
   Tests now verify that name-based struct coercion works correctly in CASE expressions.

2. Updated test expectations to match actual behavior:
   - When THEN branch executes, result uses THEN branch's field order
   - When ELSE branch executes, result uses ELSE branch's field order
   - Struct coercion requires equal field counts - mismatch causes planning error

3. Added explicit test for field count mismatch case:
   - Verifies that coercing structs with different field counts (2 fields vs 1 field)
     correctly fails during type coercion with appropriate error message

The tests now pass because:
- The optimizer fix from the previous commit prevents const-folding hangs
- Name-based struct coercion in struct_coercion function handles field reordering
- Type coercion correctly rejects field count mismatches during planning

Remove duplicate struct_cast.rs module and use the existing
nested_struct::cast_struct_column implementation instead. This
eliminates code duplication and provides a single source of truth
for struct field-by-name casting logic.

Changes:
- Add public cast_struct_array_by_name wrapper in nested_struct.rs
- Update columnar_value.rs to use nested_struct::cast_struct_array_by_name
- Update scalar/mod.rs to use nested_struct::cast_struct_array_by_name
- Remove struct_cast module from lib.rs
- Delete datafusion/common/src/struct_cast.rs

Benefits:
- Single implementation to maintain and test
- Consistent behavior across all struct casting operations
- Reduced maintenance burden for future bug fixes
- Better code cohesion in nested_struct module

Implement null fast paths for struct casting and compatibility
checks to return null struct arrays for NULL-only inputs.
Allow NULL source fields during validation. Add a unit test
covering the scenario of casting a NULL struct field into a
nested struct target.

Return plan errors directly from optimizer rule failures and update
the related test expectations. Relax the SQLogicTest expectation for
struct cast mismatches to allow for either plan-error prefix
variant.

…ind 'parquet' feature

Consolidate duplicated logic in cast_struct_column by using a single
loop for both name-overlap and positional mapping cases. This change
selects the source child by either name or position and centralizes
the cast-and-error handling, improving code clarity and maintainability.

…lding to prevent optimizer hang

…to clarify name-based casting support

…by consolidated suite

…ing tests