Pair the branches of a set operation by position

[IL-William]

Patch 9 of the engine fork that collin pins (dbt-collin decision 0028): the branches of a set operation pair by position, as SQL pairs them, not by name.

A later branch of a UNION, an INTERSECT or an EXCEPT added its columns to the target under their own names. A column that a later branch names differently from the first branch was therefore lost to whichever output column had that name, or became an output of its own:

SELECT k, a AS x FROM t1
UNION ALL
SELECT m.k, m.b FROM t2 AS m

That query came out with three outputs, k, x and b, and x derived from t1.a alone.

The change

• A later branch's columns take the first branch's names, position by position, with stars expanded.
• Only the columns the branch gives the set operation's target are renamed. A derived table or a subquery inside the branch keeps its own names.
• UNION BY NAME still pairs by name.
• A nested set operation hands its first branch's names up, so a third branch pairs with the first.

Tests

• Four new tests: a top level union, three branches in a CTE with a star, BY NAME, and a derived table inside a later branch. The first two fail without the patch.
• cte_self_join_alias_columns_do_not_leak_across_union_scopes asserted the old pairing: it expected the second branch's alias, dept_name, as an output. It now finds one output, emp_name, read from each branch's own CTE. If the second branch's alias leaked to the first CTE, it would see emp alone.
• The recursive_cte snapshot gains the edge its second branch writes: SELECT e.employee_id, e.employee_id, e.name feeds manager_id from employee_id.
• 2962 tests pass, and cargo fmt --check is clean.

Measured through collin

Project	Effect
3507 model Snowflake project, both of its compiles	cache and dumps identical
Cal-ITP (BigQuery)	3 models go from contradicted by the warehouse to confirmed; 12 phantom columns gone (uace_code, uza_name, and col_0 to col_4 twice); 4 edges added, 3 filter reads placed
Shopify, Jaffle Shop	identical
lineage-library	the one line it marked as known missing is now given

I read each moved edge against the SQL.

Nothing goes to pondpilot/flowscope.

🤖 Generated with Claude Code

https://claude.ai/code/session_017VTEw2dNNL9Pn57Hkeeirn