test: real multi-node docker topology — per-node execution, drift, and plan never live-tested across nodes

Symptom

docker-compose runs a single ClickHouse node + one Keeper. The three clusters defined in docker/clickhouse/config.d/default.xml (posthog, posthog_single_shard, posthog_migrations) are each one shard / one replica pointing at the same host.

Impact

The per-node execution model (each statement run on every node individually) is never exercised live against an actual multi-node topology: Replicated table behavior across replicas, per-shard macros, drift detection between real nodes, and plan's cross-role ordering (docs/plans/2026-06-25-multi-node-cross-role-plan.md Mode B) all lack live coverage.

Fix direction

Add a 2-node compose profile (2 replicas, or 2 shards x 1 replica) with distinct macros; gate a small live suite on it (e.g. -clickhouse -cluster) covering: per-node DDL execution reaching both nodes, dump-cluster across both, drift between them, and a plan run against the dumped topology.

Found in the 2026-07-02 deep-dive audit (docs/plans/2026-07-02-deep-dive-improvement-areas.md, E1).

Start with docker/clickhouse/config.d/default.xml and the existing Docker Compose setup, then read docs/plans/2026-06-25-multi-node-cross-role-plan.md for the required topology. Locate the live suite entry point and its -clickhouse -cluster gating. Done means a two-node profile runs coverage for per-node DDL, dump-cluster, drift, and plan against the dumped topology.