fix(optimizer): keep columns a recursive CTE reads from itself [CLAUDE]

[isaka1022]

Fixes optimizer.optimize crashing (or silently producing invalid SQL) on a self-referencing recursive CTE when the enclosing query doesn't select every CTE column.

Repro on current main:

from sqlglot import parse_one
from sqlglot.optimizer import optimize

sql = """
WITH RECURSIVE t AS (
  SELECT id, link FROM graph WHERE id = 1
  UNION ALL
  SELECT g.id, g.link FROM graph AS g, t WHERE g.id = t.link
)
SELECT id FROM t
"""
optimize(parse_one(sql, read="postgres"), schema={"graph": {"id": "INT", "link": "INT"}}, dialect="postgres")
# KeyError: 'link'  (merge_subqueries.py)

Root cause: _traverse_ctes gives the CTE body a throwaway branch scope to resolve the recursive term's self-reference. pushdown_projections records the self-reference's column usage (t.link above) against that throwaway scope, so it is lost, and the projections the recursive term still joins on get pruned from the CTE. merge_subqueries then crashes on the missing projection; depending on the query the pruning can also survive silently as invalid SQL. Remapping the usage to the real scope wouldn't help on its own, because the reversed-postorder pass visits the union before the recursive term, making the self-reference a back-edge.

Fix: treat self-referencing recursive CTEs conservatively and skip pruning their projections, mirroring the existing DISTINCT guard in the same loop. merge_subqueries._is_recursive and pushdown_predicates already special-case recursive CTEs the same way; pushdown_projections was the one rule missing the guard.

Added a regression test that asserts the optimized CTE keeps the column its recursive term joins on (it fails with a KeyError without the fix).

🤖 Generated with Claude Code

[georgesittas]

Thank you, LGTM, but I think we can improve this by checking what projections are actually referenced to avoid the over-conservative block.

I'll take it to the finish line myself.