Prepare every search benchmark replica before querying

Reproduced cold builds on two replicas during the original TermQuery. The comparison host builds with warnings as errors and scoped formatting passes.

Author: KSemenenko

benchmarks/KeyLoad.ComparisonHost/Deployment/Connections/KeyLoadDatabase.cs

@@ -36,6 +36,8 @@ internal KeyLoadDatabase(BenchmarkConnection connection)
     /// <summary>The next node's client; calls spread evenly over the cluster.</summary>
     internal KeyLoadClient Client => clients[(int)((uint)Interlocked.Increment(ref next) % (uint)clients.Length)];
 
+    internal IReadOnlyList<KeyLoadClient> NodeClients => clients;
+
     internal EntityRef Entity(string collection, string id) => new(Partition, collection, id);
 
     /// <summary>Creates a resource in the job's database before any data is written.</summary>

benchmarks/KeyLoad.ComparisonHost/Search/Drivers/KeyLoadSearchDriver.cs

@@ -32,12 +32,15 @@ public async Task InsertAsync(IReadOnlyList<SearchDocument> documents, Cancellat
         Observe((await database.CommitAsync(mutations, token)).Token);
     }
 
-    /// <summary>Returns once the lexical index covers every acknowledged insert.</summary>
+    /// <summary>Waits for the lexical index on each node that will serve measured queries.</summary>
     public async Task BuildIndexAsync(CancellationToken token)
     {
         var minimum = Volatile.Read(ref latest) ?? throw new InvalidOperationException(NothingInserted);
         var request = new WaitForIndexRequest(database.Partition, Collection, TextField, minimum);
-        _ = KeyLoadDatabase.Require(await database.Client.WaitForIndexAsync(request, token), nameof(BuildIndexAsync));
+        foreach (var client in database.NodeClients)
+        {
+            _ = KeyLoadDatabase.Require(await client.WaitForIndexAsync(request, token), nameof(BuildIndexAsync));
+        }
     }
 
     public async Task<IReadOnlyList<string>> SearchAsync(IReadOnlyList<string> terms, int limit, CancellationToken token)

docs/Features/BenchmarkComparisons.md

@@ -85,9 +85,10 @@ flowchart LR
 
 Each database and node count in this table is one job: 45 jobs. PostgreSQL runs
 as one deployment name, `PostgreSQL + pgvector`, for both relational and vector
-jobs. A load case counts records (vertices plus edges for the graph); the index
-case is one call that makes the loaded search dataset searchable; every other case
-counts requests.
+jobs. A load case counts records (vertices plus edges for the graph). The index
+case counts one operation to make the loaded dataset searchable across the
+deployment. KeyLoad waits for its index on every node used by the query driver.
+Every other case counts requests.
 
 ## Shared modules: change a dataset or scenario once