fix: don't drop the tail of a multi-hop exists path with an unanchored predicate

[barnabasJ]

The bug

An exists aggregate or exists/2 expression over a multi-hop relationship path wrongly returns true when there are zero matching related rows, whenever the predicate contains no refs:

• unfiltered exists aggregates: exists :has_policy?, [:service_line, :policies]
• constant predicates: expr(exists(service_line.policies, true))

Joins inside the exists subquery are derived from the refs in the predicate (and are left joins). A predicate with no refs never joins the remaining path, so the query degrades to "the first relationship exists":

-- exists(bucket.items, true), before:
exists((SELECT 1 FROM buckets st0 WHERE outer.bucket_id = st0.id))  -- items never queried

Relatedly, a null-satisfiable predicate (e.g. exists(bucket.items, is_nil(name))) is satisfied by the null-extended rows of the left joins.

The fix

Anchor the subquery filter with not is_nil(primary_key) at every hop of the remaining path. The anchors are a no-op for real rows, force the joins to materialize, and exclude null-extended rows:

-- exists(bucket.items, true), after:
exists((SELECT 1 FROM buckets st0
        LEFT JOIN items st1 ON st0.id = st1.bucket_id
        WHERE st1.id IS NOT NULL AND outer.bucket_id = st0.id))

Anchoring only the last hop would not be enough: no_attributes? relationships join with on: true, so their rows join even when an earlier hop of the chain is null-extended.

Tests

Regression tests live in the companion ash_postgres PR (ash-project/ash_postgres#843), since exercising this end-to-end needs a real data layer: zero-row multi-hop exists, three-hop chains broken at each level, a no_attributes? hop after a broken middle hop, and predicate scoping. The full ash_postgres suite passes against this branch (ASH_SQL_VERSION=local), including all pre-existing exists/aggregate tests.

🤖 Generated with Claude Code

[zachdaniel]

🚀 Thank you for your contribution! 🚀