feat(context): reasoning-thread continuity, L3 doc filtering, and gate hardening

Threads reasoning_content across inference calls and journal renders, filters
markdown noise out of L3 repo-knowledge retrieval, adds PlanLinter H3 checks,
and tightens filesystem tool output/dir-listing behavior surfaced by prior
live QA (see project_readloop_campaign memory).

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

@@ -142,9 +142,15 @@ class LlamaCppInferenceProvider(
 
         val baseMessages = PromptRenderer.render(request.contextPack)
         val messages = baseMessages.map { msg ->
+            val toolCalls = if (msg.role == "assistant") {
+                parseAssistantToolCalls(msg.content)
+            } else emptyList()
             LlamaCppChatMessage(
                 role = msg.role,
-                content = msg.content,
+                content = if (toolCalls.isNotEmpty()) null else msg.content,
+                reasoningContent = msg.reasoning,
+                toolCalls = toolCalls,
+                toolCallId = msg.toolCallId,
             )
         }
 
@@ -230,5 +236,40 @@ class LlamaCppInferenceProvider(
         ProviderHealth.Unavailable("Health check failed: ${e.message}")
     }
 
+    /**
+     * Parses an assistant message's content back into [ToolCallRequest] objects.
+     *
+     * The context builder's FORMAT_COMPRESS stage converts raw ToolCallRequest JSON into a compact
+     * dotted-line format (id = ..., function.name = ...). If compression is disabled, the content
+     * is still valid JSON. This handles both forms so the Gemma4 Jinja template sees the native
+     * `tool_calls` array — which it requires to render `reasoning_content` as a `<|channel>thought`
+     * block (without `tool_calls` the reasoning field is silently dropped and the model re-enters
+     * thought mode from scratch, causing the degenerate `<|channel>thought`~16384-token loop).
+     */
+    private fun parseAssistantToolCalls(content: String): List<ToolCallRequest> {
+        // Try JSON parse first (FORMAT_COMPRESS disabled — raw ToolCallRequest JSON).
+        if (content.startsWith("{\"id\"")) {
+            runCatching {
+                return listOf(json.decodeFromString<ToolCallRequest>(content))
+            }
+        }
+        // Try dotted-line format (FORMAT_COMPRESS enabled).
+        if (!content.contains("function.name")) return emptyList()
+        val lines = content.split('\n')
+        var id: String? = null
+        var name: String? = null
+        var arguments: String? = null
+        for (line in lines) {
+            when {
+                line.startsWith("id = ") -> id = line.removePrefix("id = ").trim()
+                line.startsWith("function.name = ") -> name = line.removePrefix("function.name = ").trim()
+                line.startsWith("function.arguments = ") -> arguments = line.removePrefix("function.arguments = ").trim()
+            }
+        }
+        return if (name != null && arguments != null) {
+            listOf(ToolCallRequest(id = id, function = ToolCallFunction(name = name, arguments = arguments)))
+        } else emptyList()
+    }
+
     override fun capabilities(): Set<CapabilityScore> = descriptor.capabilities
 }

infrastructure/inference/openai_compat/src/main/kotlin/com/correx/infrastructure/inference/openai/OpenAiCompatInferenceProvider.kt

@@ -88,7 +88,7 @@ class OpenAiCompatInferenceProvider(
             // tool_calls turn (we don't carry tool_call_ids through the context pack). Fold any
             // non-standard role into a user turn so the model still sees the content.
             if (msg.role in STANDARD_ROLES) {
-                OpenAiChatMessage(role = msg.role, content = msg.content)
+                OpenAiChatMessage(role = msg.role, content = msg.content, reasoningContent = msg.reasoning)
             } else {
                 OpenAiChatMessage(role = "user", content = "[${msg.role}] ${msg.content}")
             }