[IL-William]

Patch 12 of the engine fork that collin pins (dbt-collin decision 0028). A scalar subquery in a select list is now analyzed, and its column feeds the item that reads it.

Until now it never was: only the item's column references were read, and those skip a subquery. With no source, the output column got a relation-level edge from the FROM table, carrying the subquery's text:

SELECT b.book_id,
       (SELECT max(r.stars) FROM reviews AS r WHERE r.book_id = b.book_id) AS best_stars,
       (SELECT max(r.written_at) FROM reviews AS r) AS last_review_at
FROM books AS b

came out with best_stars and last_review_at fed by the books table, and reviews nowhere. A reader of that edge's text could only guess, and collin's credited b.book_id to best_stars.

The change

• Each Expr::Subquery an item reads as a value, alone or inside a function, a CASE, a cast or an operator, is analyzed into a (subquery) node of its own, through the helper analyze_predicate_subquery now shares, and its columns are added to the item's sources.
• Its columns are registered in the enclosing select's scope under the subquery's own key (register_scalar_subquery_in_scope), until that select ends, so a later item reading the item's alias reaches them too. The key is a subquery alias, never a relation of the select: unqualified names, wildcards, items with no source (count(*), a literal) and the implied schema are as before.
• The enclosing select's grouping (grouping_columns, has_group_by) is saved and restored around any subquery it analyzes; the subquery's own select used to clear it, so later items lost their grouping keys.
• The walk follows the same value positions as collect_column_refs, lists every variant as it does, and says they must stay in step. Row-testing forms are not walked into: EXISTS, IN (subquery), and the subquery of = ANY (...) or = ALL (...) in a select list are left as they were.

Tests

• correlated and uncorrelated subqueries feed their columns from reviews, with no edge from books;
• inside COALESCE, a CASE, a cast and a concatenation, beside the outer columns read;
• inside TRIM, SUBSTRING ... FROM ... FOR and AT TIME ZONE;
• in a grouped select, the later items keep their grouping keys (fails without the restore);
• in a CTE body;
• the correlated subquery's own filter stays on reviews, and b resolves;
• EXISTS in a select list unchanged.

Three of the new tests fail without the patch. cargo test -p flowscope-core passes, cargo fmt --all -- --check and CI's cargo clippy --workspace --locked -- -D warnings are clean. The same commit applies unchanged to collin and master.

Measured through collin

With a local [patch] on this branch:

Project	Effect
lineage-library	the four book_ratings value lines collin marks now hold: best_stars <- raw.reviews.stars aggregate and last_review_at <- raw.reviews.written_at aggregate given, best_stars <- books.book_id and the lost last_review_at gone; nothing else moves
Jaffle Shop	cache identical, --indirect too
Shopify	cache identical, --indirect too

🤖 Generated with Claude Code

https://claude.ai/code/session_018LDZwk297SKdJXXLpTFZEh