GeezerVison

Review: the fix looks correct. It deep-copies the result of g and gives every nested closure fresh parameters, so the post-processing and the empty-set result no longer share lambda parameters.

What I ran

Missing tests (medium): The PR has no test for nested lambdas or for several fused columns. These are the cases where shared parameters are most likely to come back. Here is a test you could add:

@Test
public void fusedNestedLambdasForGen2() {
    var nested = this.compileForGen2("""
            CREATE TABLE t(k INT, c INT);
            CREATE VIEW v AS SELECT k,
              ARRAY_MAX(TRANSFORM(TRANSFORM(ARRAY_AGG(c), x -> x * 2), y -> y + 1)) AS m
            FROM t GROUP BY k;""");
    Assert.assertEquals(0, storedArrays(nested));
    var two = this.compileForGen2("""
            CREATE TABLE t(k INT, c INT, d INT);
            CREATE VIEW v AS SELECT k,
              ARRAY_MAX(TRANSFORM(ARRAY_AGG(c), x -> x * 2)) AS m,
              ARRAY_MIN(TRANSFORM(ARRAY_AGG(d), x -> x + 1)) AS n
            FROM t GROUP BY k;""");
    Assert.assertEquals(0, storedArrays(two));
}

Commit message (medium): The body has only the Signed-off-by line. Please explain the why: Gen-2 keeps the aggregate list, and CanonicalForm then requires each parameter to belong to exactly one closure. Please also include the already mapped error message. The description of #7718 already explains this well and could be reused.