fix: store actual rendered prompt artifact instead of contextpack.toString()

Extract message-rendering logic from LlamaCppInferenceProvider into a
provider-agnostic PromptRenderer in core:inference. SessionOrchestrator
now serializes the rendered ChatMessage list as JSON for the prompt
artifact, so the audit trail reflects what the LLM actually saw.

Resolves the TODO at SessionOrchestrator.kt:691.

core/inference/src/main/kotlin/com/correx/core/inference/PromptRenderer.kt

@@ -0,0 +1,41 @@
+package com.correx.core.inference
+
+import com.correx.core.context.model.ContextLayer
+import com.correx.core.context.model.ContextPack
+import com.correx.core.context.model.EntryRole
+import kotlinx.serialization.Serializable
+
+@Serializable
+data class ChatMessage(
+    val role: String,
+    val content: String,
+)
+
+object PromptRenderer {
+    fun render(contextPack: ContextPack): List<ChatMessage> {
+        val sorted = contextPack.layers.entries.sortedBy { it.key.ordinal }
+        val systemContent = sorted
+            .filter { it.key == ContextLayer.L0 }
+            .flatMap { it.value }
+            .joinToString("\n\n") { it.content }
+            .takeIf { it.isNotBlank() }
+        val conversationMessages = sorted.filter { it.key != ContextLayer.L0 }.flatMap { (_, entries) ->
+            entries.map { entry ->
+                ChatMessage(
+                    role = when (entry.role) {
+                        EntryRole.SYSTEM -> "system"
+                        EntryRole.ASSISTANT -> "assistant"
+                        EntryRole.TOOL -> "tool"
+                        EntryRole.USER -> "user"
+                    },
+                    content = entry.content,
+                )
+            }
+        }
+        val messages = buildList {
+            systemContent?.let { add(ChatMessage("system", it)) }
+            addAll(conversationMessages)
+        }
+        return messages.ifEmpty { listOf(ChatMessage("user", "")) }
+    }
+}

core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/SessionOrchestrator.kt

@@ -61,6 +61,7 @@ import com.correx.core.inference.InferenceRepository
 import com.correx.core.inference.InferenceRequest
 import com.correx.core.inference.InferenceResponse
 import com.correx.core.inference.InferenceRouter
+import com.correx.core.inference.PromptRenderer
 import com.correx.core.inference.ResponseFormat
 import com.correx.core.inference.Tokenizer
 import com.correx.core.inference.ToolCallRequest
@@ -91,6 +92,7 @@ import kotlinx.coroutines.TimeoutCancellationException
 import kotlinx.coroutines.withContext
 import kotlinx.coroutines.withTimeout
 import kotlinx.datetime.Clock
+import kotlinx.serialization.encodeToString
 import kotlinx.serialization.json.Json
 import kotlinx.serialization.json.JsonObject
 import org.slf4j.LoggerFactory
@@ -688,8 +690,8 @@ abstract class SessionOrchestrator(
                 ),
         )
 
-        // TODO: capture rendered prompt; currently contextpack.toString()
-        val promptArtifactId = artifactStore.put(contextPack.toString().toByteArray())
+        val renderedMessages = PromptRenderer.render(contextPack)
+        val promptArtifactId = artifactStore.put(Json.encodeToString(renderedMessages).toByteArray())
         emit(sessionId, InferenceStartedEvent(requestId, sessionId, stageId, provider.id, promptArtifactId))
 
         if (isCancelled(sessionId)) return InferenceResult.Cancelled

infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppApiModels.kt

@@ -1,5 +1,6 @@
 package com.correx.infrastructure.inference.llama.cpp
 
+import com.correx.core.inference.ChatMessage
 import com.correx.core.inference.ToolCallRequest
 import com.correx.core.inference.ToolDefinition
 import kotlinx.serialization.SerialName
@@ -8,7 +9,7 @@ import kotlinx.serialization.Serializable
 @Serializable
 data class ChatCompletionRequest(
     val model: String,
-    val messages: List<ChatMessage>,
+    val messages: List<LlamaCppChatMessage>,
     val temperature: Double,
     @SerialName("top_p") val topP: Double,
     @SerialName("max_tokens") val maxTokens: Int,
@@ -20,7 +21,7 @@ data class ChatCompletionRequest(
 )
 
 @Serializable
-data class ChatMessage(
+data class LlamaCppChatMessage(
     val role: String,
     val content: String?,
     @SerialName("tool_calls") val toolCalls: List<ToolCallRequest> = emptyList(),
@@ -36,7 +37,7 @@ data class ChatCompletionResponse(
 
 @Serializable
 data class Choice(
-    val message: ChatMessage,
+    val message: LlamaCppChatMessage,
     @SerialName("finish_reason") val finishReason: String,
 )
 

infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppInferenceProvider.kt

@@ -1,14 +1,12 @@
 package com.correx.infrastructure.inference.llama.cpp
 
-import com.correx.core.context.model.ContextLayer
-import com.correx.core.context.model.ContextPack
-import com.correx.core.context.model.EntryRole
 import com.correx.core.events.types.ProviderId
 import com.correx.core.inference.CapabilityScore
 import com.correx.core.inference.FinishReason
 import com.correx.core.inference.InferenceProvider
 import com.correx.core.inference.InferenceRequest
 import com.correx.core.inference.InferenceResponse
+import com.correx.core.inference.PromptRenderer
 import com.correx.core.inference.ProviderHealth
 import com.correx.core.inference.ResponseFormat
 import com.correx.core.inference.TokenUsage
@@ -79,7 +77,13 @@ class LlamaCppInferenceProvider(
     override suspend fun infer(request: InferenceRequest): InferenceResponse {
         val startTime = System.currentTimeMillis()
 
-        val messages = buildMessages(request.contextPack)
+        val baseMessages = PromptRenderer.render(request.contextPack)
+        val messages = baseMessages.map { msg ->
+            LlamaCppChatMessage(
+                role = msg.role,
+                content = msg.content,
+            )
+        }
 
         val grammar = when (val fmt = request.responseFormat) {
             is ResponseFormat.Json -> GbnfGrammarConverter.convert(fmt.schema)
@@ -142,31 +146,4 @@ class LlamaCppInferenceProvider(
     }
 
     override fun capabilities(): Set<CapabilityScore> = descriptor.capabilities
-
-    private fun buildMessages(contextPack: ContextPack): List<ChatMessage> {
-        val sorted = contextPack.layers.entries.sortedBy { it.key.ordinal }
-        val systemContent = sorted
-            .filter { it.key == ContextLayer.L0 }
-            .flatMap { it.value }
-            .joinToString("\n\n") { it.content }
-            .takeIf { it.isNotBlank() }
-        val conversationMessages = sorted.filter { it.key != ContextLayer.L0 }.flatMap { (_, entries) ->
-            entries.map { entry ->
-                ChatMessage(
-                    role = when (entry.role) {
-                        EntryRole.SYSTEM -> "system"
-                        EntryRole.ASSISTANT -> "assistant"
-                        EntryRole.TOOL -> "tool"
-                        EntryRole.USER -> "user"
-                    },
-                    content = entry.content,
-                )
-            }
-        }
-        val messages = buildList {
-            systemContent?.let { add(ChatMessage("system", it)) }
-            addAll(conversationMessages)
-        }
-        return messages.ifEmpty { listOf(ChatMessage("user", "")) }
-    }
 }