feat(server): wire EmbeddingRelevanceScorer into context builder, default level 4

Move embedder construction ahead of the engines so the context pack builder ranks
freeform turns by query relevance (query-conditioned selection). Default
compression_level 4 (format + static cache + token pruning + age-tiering) for a 12b
on a 16k window.

apps/server/src/main/kotlin/com/correx/apps/server/Main.kt

@@ -18,6 +18,7 @@ import com.correx.core.config.ProfileLoader
 import com.correx.core.config.ProviderConfig
 import com.correx.core.context.builder.DefaultContextPackBuilder
 import com.correx.core.context.compression.CompressionPolicy
+import com.correx.core.inference.EmbeddingRelevanceScorer
 import com.correx.infrastructure.compression.HttpTokenPruner
 import com.correx.core.context.compression.DefaultContextCompressor
 import com.correx.core.context.model.TokenBudget
@@ -291,12 +292,16 @@ fun main() {
         WorkspaceTools(registry = wsRegistry, executor = wsExecutor)
     }
 
+    // Constructed before the engines so the context builder can rank freeform turns by relevance
+    // to the current query (pipeline §6, query-conditioned selection). Reused by L3 retrieval below.
+    val embedder = InfrastructureModule.createEmbedderFromConfig(correxConfig.router.embedder)
     val engines = OrchestratorEngines(
         transitionResolver = DefaultTransitionResolver { condition, ctx -> condition.evaluate(ctx) },
         contextPackBuilder = DefaultContextPackBuilder(
             DefaultContextCompressor(),
             policy = CompressionPolicy(correxConfig.orchestration.compressionLevel),
             tokenPruner = HttpTokenPruner(baseUrl = correxConfig.orchestration.tokenPrunerUrl),
+            relevanceScorer = EmbeddingRelevanceScorer(embedder),
         ),
         inferenceRouter = inferenceRouter,
         validationPipeline = ValidationPipeline(
@@ -340,7 +345,6 @@ fun main() {
     val artifactKindRegistry = DefaultArtifactKindRegistry().also { reg ->
         configArtifactKindsEarly.forEach { reg.register(it) }
     }
-    val embedder = InfrastructureModule.createEmbedderFromConfig(correxConfig.router.embedder)
     val l3MemoryStore = InfrastructureModule.createL3MemoryStoreFromConfig(correxConfig.router.l3)
     // Rebuild the L3 metadata map from the ChatTurnEvent log before serving queries:
     // persisted vectors survive restart but the metadata map does not (no-op for

core/config/src/main/kotlin/com/correx/core/config/ConfigLoader.kt

@@ -624,7 +624,7 @@ object ConfigLoader {
                 asInt(orchestrationSection["journal_compaction_token_threshold"], 2_000),
             resumeAbandonedMaxAgeMinutes =
                 asLong(orchestrationSection["resume_abandoned_max_age_minutes"], 1_440),
-            compressionLevel = asInt(orchestrationSection["compression_level"], 2),
+            compressionLevel = asInt(orchestrationSection["compression_level"], 4),
             tokenPrunerUrl =
                 asString(orchestrationSection["token_pruner_url"], "http://127.0.0.1:8199"),
         )

core/config/src/main/kotlin/com/correx/core/config/CorrexConfig.kt

@@ -64,9 +64,10 @@ data class OrchestrationKnobs(
     /**
      * Context compression pipeline level (docs/plans/correx-compression-pipeline.md §5), additive
      * 1..9. Default 2 = free format-compress + static cache. Raise toward the 16k wall; ≥3 needs
-     * the LLMLingua-2 sidecar at [tokenPrunerUrl] (fails open if absent).
+     * the LLMLingua-2 sidecar at [tokenPrunerUrl] (fails open if absent). Default 4 = format +
+     * static cache + token pruning + age-tiering, sized for a 12b on a 16k window.
      */
-    val compressionLevel: Int = 2,
+    val compressionLevel: Int = 4,
     val tokenPrunerUrl: String = "http://127.0.0.1:8199",
 )