GeezerVison

A scalar subquery yields NULL when it matches no row, but the legacy analyzer gave its column the nullability of the column it selects, so (select r.id from reviews r where r.book_id = b.id limit 1) came out non-null. This makes it nullable unless the subquery always yields exactly one row: an aggregate with no GROUP BY, HAVING or LIMIT ((select count(*) …) stays non-null), or a SELECT with no FROM.

Fixes #4262. The new nullable_scalar_subquery fixture covers a subquery through a view and inline; it fails without the change in both the base and managed-db contexts. With PostgreSQL available, every TestReplay case passes, and no existing expected output changes.

CASE … ELSE NULL::type has the same problem on this path; I've left it to #3739, which reworks CASE (see my comment there about column-reference results).

🤖 Generated with Claude Code