Recursive CTE silently casts recursive term to the static term's type, producing wrong results or a non-terminating query #25893

Describe the bug

A recursive CTE silently casts the recursive term's output to the static term's types, even when the cast loses data (for example Float64 → Int64). The query returns wrong values. If the stop condition depends on the lost part, the query never terminates.

LogicalPlanBuilder::to_recursive_query calls coerce_plan_expr_for_schema(recursive_term, static_schema), which casts to any castable type (added for #9794).

To Reproduce

Tested on main (1d9be2e), default dialect, datafusion-cli:

-- Wrong result: expected n = 1, 1.5, 2, 2.5
WITH RECURSIVE r AS (
  SELECT 1 AS n, 0 AS i
  UNION ALL
  SELECT n + 0.5, i + 1 FROM r WHERE i < 3
) SELECT n, arrow_typeof(n) FROM r ORDER BY i;
-- 1 Int64 / 1 Int64 / 1 Int64 / 1 Int64

-- Never terminates: n is truncated from 1.5 back to 1, so n < 2 stays true
WITH RECURSIVE r AS (
  SELECT 1 AS n
  UNION ALL
  SELECT n + 0.5 FROM r WHERE n < 2
) SELECT count(*) FROM r;

EXPLAIN shows the recursive term as CAST(CAST(n AS Float64) + 0.5 AS Int64).

Expected behavior

An error at planning time, as PostgreSQL does:

ERROR:  recursive query "r" column 1 has type integer in non-recursive term but type numeric overall
HINT:  Cast the output of the non-recursive term to the correct type.

Additional context

Suggested fix: in to_recursive_query, return a planning error when the cast from the recursive type to the static type can lose data — float/decimal → integer, and changes of type category such as Utf8 → Int64. Keep the current cast for integer → integer so the existing #9794 test (a::bigint + 2 into Int32) still passes. The error message should tell the user to cast the non-recursive term.

Alternative: widen both terms to their common type, as UNION does. That requires planning the recursive term again against the widened work-table schema.