Add Queryable nodes for the new LINQ operators

[davidanthoff]

Second of four PRs adding the LINQ operators Query.jl is missing. Depends on QueryOperators.jl#56.

CI will stay red until QueryOperators 1.2.0 is registered. The [compat] bound is an ordinary QueryOperators = "1.2"; no feature guards. Locally the full suite passes (22 items, 14 new) against the QueryOperators branch.

What it does

Adds a Queryable node for each operator introduced in QueryOperators 1.2, so a backend sees them in the query tree instead of having the in-memory implementation run behind its back — outer joins, the seven set operators, the four partitioning operators, reverse/shuffle/index, append/prepend/zip, count_by/aggregate_by/chunk, and of_type/cast. Each gets a describe_node method so @queryplan() renders it rather than falling through to the type-name fallback (a test asserts that).

Three things worth reviewing

A QueryableBinary abstract type. get_source and _collect_nodes! needed a method per node that combines two sources, since those carry outer/inner rather than source. Outer joins, set operators and zip would have added eleven more pairs of near-identical methods. QueryableJoin and QueryableGroupJoin are now subtypes of QueryableBinary, the tree walk has one method covering every two-input node, and a new node only has to pick the right supertype. QueryableBinary <: Queryable and the concrete types are unchanged, so node isa QueryableJoin and anything dispatching on Queryable behave exactly as before — QueryDuckDB needs no change for this.

Terminal-operator support, and a bug it fixes. The tree had no notion of an operator that returns a value, and QueryOperators.count had no method taking a Queryable at all — so df |> @duckdb() |> @count() failed outright today. QueryableScalar now captures one, and execute_scalar dispatches on the root source type:

execute_scalar(q::QueryableScalar) = _execute_scalar(get_source(q.source), q)

_execute_scalar(::QueryableSource, q::QueryableScalar) = q.f(_materialize(q.source), q.args...)

A backend specialises _execute_scalar on its own source type to push the operator down; the fallback materialises and delegates to the in-memory implementation, so every backend is correct for free and push-down is a pure optimisation. count is fixed as a side effect and a regression test pins it. QueryDuckDB will specialise this in the fourth PR.

order adds no node. It lowers to orderby with an identity key selector, so thenby can still follow it and a backend can recognise the identity selector via the is_identity_lambda helper QueryDuckDB already has, and emit ORDER BY ALL.

Smaller: the plain and _by forms of the set operators share a node, with a nothing key selector marking the plain form; zip does the same with its result selector.

Tests call QueryOperators.* directly rather than through Query macros, since those land in the Query.jl PR.

🤖 Generated with Claude Code