GeezerVison

What I ran: ./build.sh, then mvn test -pl SQL-compiler for MultisetAnalysisTests, OtherTests, MetadataTests, EquivalenceTests, ShareIndexesTests, TopKUnusedFieldsTests and DBSPWindowOperatorTests. All 159 tests passed (1 skipped).

QA tests I added locally (all pass, including the Rust run):

missing (medium): test coverage. MultisetAnalysis has about 60 operator rules, but the PR tests only 4 of them directly: index over a keyed table, deindex with an empty key, SAFE_CAST, and recursion. There are no direct tests for the rules that now drive mark_distinct() in generated code:

• join / left join / star join with map
• Lag, Rank, RowNumber, IndexedTopK
• Hop / Window
• ChainOperator
• unsafe narrowing integer casts, which are treated as injective because they raise an error on overflow

If one of these rules is wrong, mark_distinct() lands on a multiset, and the runtime can silently return wrong results. That is worth an end-to-end stepWeightOne test for each rule. The two QA tests above can serve as templates:

@Test
public void qaProjectionOverSet() {
    var ccs = this.getCCS("""
            CREATE TABLE t (a INT, b INT);
            CREATE VIEW v AS SELECT DISTINCT a FROM (SELECT DISTINCT a, b FROM t);""");
    ccs.stepWeightOne("INSERT INTO t VALUES (1, 1), (1, 2);", """
             a
            ---
             1""");
}

missing (medium): breaking change not called out. Persistent ids change for existing programs; details are in the inline comment. The PR description doesn't mention it, nor does it describe any manual testing.