An Int64 cast of an Int64 field's address keeps a garbage high word

Found by the Pascal cross-target differential: test_record_layout_stress failed packed-record offset rows (8, 16, 29) on arm32 and riscv32, where no ABI could move a packed field. The layout was right; the offset arithmetic was not.

Resolution (2026-09-28)

@r.i lowers to an IR_FIELD tagged with the FIELD's kind (the pointee, which IRValueKind documents). The AN_PTR_CAST lowering asked IRTk whether the operand was "already 64 bits", got Int64, and skipped the widen; the 32-bit pair path then read the address node as an Int64 pair. The widen test now asks IRValueKind, and inside the widen the freshly lowered address operand is retagged tyPointer (as IRLowerDestAddress already does) so the pair path zero-extends it.

Fixture test/test_an_int64_cast_of_a_fields_address_widens_the_pointer.pas (.expected by FPC 3.2.2), in test-core native and cross. The remaining test_record_layout_stress rows on i386 are the psABI's 4-byte member alignment of Int64/Double (TypeFieldAlign, deliberate), not this.