Column transform keeps only the last hop's classification

[funcpp]

Found while reviewing #13 (#25). #25 fixed the case where a column reaches no source; this is the case where it reaches one through an intermediate column, and loses the classification on the way.

A column is classified by its last hop only

SELECT c FROM (SELECT SUM(x) AS c FROM t) d

c  ←  t.x    (direct)

t.x is carried across the derived table correctly. The aggregation is not — c is reported as a plain pass-through of t.x. The same holds through a CTE, and whether or not the inner projection has sources:

WITH x AS (SELECT COUNT(*) AS c FROM t) SELECT c FROM x   -- Direct
SELECT c FROM (SELECT COUNT(*) AS c FROM t) d             -- Direct
SELECT c FROM (SELECT SUM(x) AS c FROM t) d               -- Direct, t.x carried

Cause

collect_output_sources walks transitively to the leaves but records only the kind of the immediate incoming edge:

for &edge_idx in &incoming[node_id] {
    let edge = &graph.edges[edge_idx];
    let (sub_sources, sub_back) =
        collect_leaf_origins(edge.from, graph, resolved, incoming, visited);
    for _ in &sub_sources {
        kinds.push(edge.kind.clone());   // the immediate edge, every time
    }
    ...
}

collect_leaf_origins returns origins and nothing else, so every kind met deeper in the walk is discarded. For SELECT c FROM (SELECT SUM(x) AS c FROM t) d the immediate edge is Ref(c) → Output(c) with kind Direct, because the outer projection really is a plain reference; the inner Ident(x) → Output(c) edge carrying ViaAggregation is never consulted.

The column's own kind, added in #25, does not reach this either. It describes the outer projection, which is correctly Direct.

Consequence: the star form disagrees with the named form

Expanding a star classifies each inner column from that column's own node, taking no second hop, so the same derived table answers differently depending on how the column is selected:

SELECT * FROM (SELECT COUNT(*) AS c FROM t) d   -- Aggregation
SELECT c    FROM (SELECT COUNT(*) AS c FROM t) d -- Direct

Pinned by only_the_last_hop_classifies_the_column in sqllineage/tests/cte.rs, so whichever way this is resolved, the change is visible.

What a fix needs

collect_leaf_origins has to report the kinds it traversed, and collect_output_sources has to merge them rather than stamping the immediate edge's kind per source. Worth settling at the same time: whether an aggregate two hops down should surface at all, or whether Direct is the honest answer for a projection that is genuinely a plain reference to a column that happens to have been aggregated upstream. The tests assume the former.

This is in the area #14 reworks, so it may be resolved there.