major refactor

detect/graph.py

@@ -326,6 +326,7 @@ def node_compile_report(state: DetectState) -> dict:
 
 _CHECKPOINT_ENABLED = os.environ.get("MPR_CHECKPOINT", "").strip() == "1"
 _frames_manifest: dict[str, dict[int, str]] = {}  # job_id → manifest (cached per job)
+_latest_checkpoint: dict[str, str] = {}  # job_id → latest checkpoint_id
 
 
 class PipelineCancelled(Exception):
@@ -361,17 +362,33 @@ def _checkpointing_node(node_name: str, node_fn):
         if not job_id:
             return result
 
-        from detect.checkpoint import save_checkpoint, save_frames
+        from detect.checkpoint import save_stage_output, save_frames
+        from detect.stages.base import _REGISTRY
 
         merged = {**state, **result}
 
-        # Save frames once (first checkpoint), reuse manifest after
+        # Save frames once (first node), reuse manifest after
         manifest = _frames_manifest.get(job_id)
         if manifest is None and node_name == "extract_frames":
             manifest = save_frames(job_id, merged.get("frames", []))
             _frames_manifest[job_id] = manifest
 
-        save_checkpoint(job_id, node_name, stage_index, merged, frames_manifest=manifest)
+        # Serialize stage output using the stage's serialize_fn if available
+        stage_cls = _REGISTRY.get(node_name)
+        serialize_fn = getattr(getattr(stage_cls, "definition", None), "serialize_fn", None)
+        if serialize_fn:
+            output_json = serialize_fn(merged, job_id)
+        else:
+            output_json = {}
+
+        parent_id = _latest_checkpoint.get(job_id)
+        new_checkpoint_id = save_stage_output(
+            timeline_id=job_id,
+            parent_checkpoint_id=parent_id,
+            stage_name=node_name,
+            output_json=output_json,
+        )
+        _latest_checkpoint[job_id] = new_checkpoint_id
         return result
 
     wrapper.__name__ = node_fn.__name__