feat(compressors): refactor pipeline with FramePacket dataclass and unified process_batch

- Add FramePacket dataclass to encapsulate per-image pipeline state
- Rename internal methods with underscore prefix convention
- Replace separate filter_batch/crop_batch with unified process_batch method
- Update notebook to use new HashPipeline API

notebooks/verification.py

@@ -160,7 +160,6 @@ def collect_views(
 
 @app.cell
 def build_scene_graph(
-    Image,
     ObjectNode,
     SimpleSceneGraph,
     cfg_manager,
@@ -168,9 +167,7 @@ def build_scene_graph(
     pipeline,
     room_nodes,
     room_view_dataset,
-    torch,
 ):
-    """Build scene graph using step-by-step pipeline to capture cropped images."""
     scene_graph = SimpleSceneGraph(
         rooms={_room.room_id: _room for _room in room_nodes},
         objects={},
@@ -185,36 +182,10 @@ def build_scene_graph(
     _images = [item[2] for item in room_view_dataset]
     _metadata = [(item[0], item[1]) for item in room_view_dataset]
 
-    # Step 1: Detect objects.
     _text_labels = ["object"]
-    _detections = pipeline.detect_batch(_images, _text_labels)
-
-    # Step 2: Segment with SAM.
-    _bboxes_per_image = [[_d["bbox"] for _d in _dets] for _dets in _detections]
-    _masks = pipeline.segment_batch(_images, _bboxes_per_image)
-
-    # Step 3: Filter masks.
-    _filtered = pipeline.filter_batch(_images, _masks)
-
-    # Step 4: Crop images.
-    _cropped_images = pipeline.crop_batch(_filtered, _masks, _detections)
-
-    # Step 5: Extract DINO features and compress to hash.
-    _batch_size = 32
-    _all_bits = []
-    for _i in range(0, len(_cropped_images), _batch_size):
-        _batch = _cropped_images[_i : _i + _batch_size]
-        _tokens = pipeline.extract_dino_batch(_batch)
-        _bits = pipeline.compress_batch(_tokens)
-        _all_bits.append(_bits)
-
-    hash_tensor = (
-        torch.cat(_all_bits, dim=0)
-        if _all_bits
-        else torch.empty(
-            (0, pipeline.hash_bits), dtype=torch.int32, device=pipeline.device
-        )
-    )
+    _output = pipeline.process_batch(_images, _text_labels, batch_size=32)
+    _cropped_images = _output.cropped_images
+    hash_tensor = _output.hash_bits
 
     # Step 6: Create ObjectNodes and save cropped images.
     for _idx, (_cropped, _hash_bits) in enumerate(zip(_cropped_images, hash_tensor)):
@@ -238,8 +209,13 @@ def build_scene_graph(
             last_seen_frame=_view_idx,
         )
 
+    _fallback_count = sum(
+        1 for _meta in _output.debug_meta if _meta["fallback_reason"] is not None
+    )
+
     print(f"Created {len(scene_graph.objects)} objects")
     print(f"Saved cropped images to: {output_dir}")
+    print(f"Fallback frames: {_fallback_count}/{len(_output.debug_meta)}")
 
     return hash_tensor, object_images, output_dir, scene_graph
 
@@ -302,19 +278,12 @@ def query_matching(
     if file_upload.value:
         _query_image = Image.open(io.BytesIO(file_upload.contents())).convert("RGB")
 
-        # Step-by-step processing to get cropped query image.
         _text_labels = ["object"]
-        _detections = pipeline.detect_batch([_query_image], _text_labels)
-        _bboxes = [[_d["bbox"] for _d in _dets] for _dets in _detections]
-        _masks = pipeline.segment_batch([_query_image], _bboxes)
-        _filtered = pipeline.filter_batch([_query_image], _masks)
-        _cropped = pipeline.crop_batch(_filtered, _masks, _detections)
-
-        _tokens = pipeline.extract_dino_batch(_cropped)
-        _query_bits = pipeline.compress_batch(_tokens)
+        _output = pipeline.process_batch([_query_image], _text_labels, batch_size=1)
+        _query_bits = _output.hash_bits
 
         if _query_bits.numel() > 0:
-            query_cropped = _cropped[0]
+            query_cropped = _output.cropped_images[0]
             _query_tensor = _query_bits[0].int()
 
             _obj_ids = list(scene_graph.objects.keys())