Physical CastExpr display hides nullability changes, so union coercion casts look like no-ops #26358

Is your feature request related to a problem or challenge?

UnionExec aligns its inputs to the merged union schema. When an input differs only in nullability, it wraps that input in a ProjectionExec of CastExpr::new_with_target_field, which keeps the data type and changes only the field's nullability.

CastExpr's Display prints only the data type, so in EXPLAIN this cast looks like a no-op:

ProjectionExec: expr=[..., CAST(trf_id@11 AS Int32) as trf_id, ...]

trf_id is already Int32. The only thing this cast does is make the column nullable, because another union input is nullable. Nothing in the plan says that, so a reader has to go back to the source code to see why the cast exists. We hit this when we changed one table's column from nullable to NOT NULL. A new same-type cast showed up in every plan that unions that table with a nullable input.

The same thing happens in DataFusion's own plans, for example in cte.slt:

ProjectionExec: expr=[CAST(1 AS Int64) as id]

Describe the solution you'd like

When the cast has an explicit target field that is nullable, print the nullability the way schemas already do, with ;N:

ProjectionExec: expr=[CAST(1 AS Int64;N) as id]

Casts built from a bare type (CastExpr::new) are unchanged. Their nullability follows the input, so there is nothing extra to show. fmt_sql is also unchanged.

I have a small patch ready. It updates goldens in 6 slt files, 1 insta snapshot and the pruning predicate unit tests. In every one of them the only change is the added ;N.

Describe alternatives you've considered

• Rely on datafusion.explain.show_schema. That shows the projection's output schema, but it is verbose and off by default. It also doesn't tie the nullability to the cast that changed it.
• Print ;N only when the nullability actually changes. Display has no access to the input schema, so it can't tell.
• Use a different notation, for example CAST(a AS Int32 NULL). ;N matches the existing schema display, but I'm happy to change it.

Additional context

The patch also shows a case that was hard to see before. rewrite_expr_to_prunable rebuilds a plain CAST(col AS T) as CastExpr::new_with_target_field(.., cast.target_field(), None). For a plain cast, target_field() is a placeholder nullable field. So the rewrite turns a cast whose nullability follows its input into one that is always nullable, and it drops the original cast options (None). This is harmless for pruning, since statistics columns are nullable anyway. It is visible now as CAST(c1_min@0 AS Int64;N) in the pruning predicate display.