cluster-controller: drop unused node_direct join from node_drain::store #885

Problem

node_api_row in crates/libs/cluster/controller/server/src/node_drain/store.rs:545-571 queries a source node's OwnedNodeApiTarget with LEFT JOIN node_direct nd ON nd.node_id = n.id to populate direct_endpoint. Because of that join, node_drain/store.rs carries an exception in architecture-baseline.toml under [pattern_ratchet.direct_tables].

However, the node-drain reconciler is Kubernetes-only (SubstrateCapabilities::DIRECT.supports_node_drain is false), and its sole consumer of node_api_row is promotion_node_api_client (store.rs:2634), which passes target directly to self.substrate.kubernetes() (KubernetesSubstrate::resolve_node_api_endpoint). KubernetesSubstrate addresses nodes exclusively via NodeAddressing::SubstrateNodeId (nk.k8s_node_name) and never reads direct_endpoint.

Why it matters

ARCHITECTURE.toml fences node_direct to the Direct substrate (crates/libs/api/src/controllers/substrate/direct*). Keeping an unused LEFT JOIN node_direct in a Kubernetes-only reconciler pays an unnecessary join on every uncached source-node lookup during drain and keeps node_drain/store.rs in [pattern_ratchet.direct_tables].

Proposed approach

1. Remove LEFT JOIN node_direct nd ON nd.node_id = n.id from node_api_row in crates/libs/cluster/controller/server/src/node_drain/store.rs and set direct_endpoint: None on the returned OwnedNodeApiTarget.
2. Remove "crates/libs/cluster/controller/server/src/node_drain/store.rs" = 1 from [pattern_ratchet.direct_tables] in architecture-baseline.toml (cargo arch --write-baseline=direct_tables).

Severity

Low (removes an unused table join and ratchets down pattern_ratchet.direct_tables by one file).