docs: require explicit backend for legacy walks

recognization/code/app.py

@@ -293,8 +293,11 @@ def create_app(
 
     scans = scans if scans is not None else build_legacy_scans(engine)
     # A live session can be recorded so a real scan becomes data to label later.
-    recorder = FrameRecorder(settings.record_frames, every=settings.record_every) \
-        if settings.record_frames else None
+    recorder = FrameRecorder(
+        settings.record_frames,
+        every=settings.record_every,
+        identity_backend=engine.settings.identity_backend,
+    ) if settings.record_frames else None
     samples = {item["id"]: item for item in get_available_shelf_images("add")}
     # path -> (mtime_ns, size, width, height). Dropped when the file changes.
     sample_dimensions: Dict[str, Tuple[int, int, int, int]] = {}
@@ -1138,7 +1141,8 @@ def main() -> None:
     engine = build_default_engine(engine_settings)
     if (engine_settings.identity_backend or "").lower() == "gallery":
         if engine.identity_gallery() is None:
-            raise RuntimeError(
+            engine.shutdown()
+            raise SystemExit(
                 "gallery identity backend requires local DINOv2 weights; refusing to start"
             )
     warm_ms = engine.warm_up()

recognization/code/eval_live.py

@@ -15,8 +15,11 @@ store was enrolled; scoring them is training-set accuracy.
 
 Prints a human count table and writes the same report as JSON (default
 <recording>/eval_live.json). Missing input exits 2 with {"error": ...} on
-stdout. --identity gallery uses the same LiveSession path as the server and
-exits 2 when that gallery cannot load; it does not score the matcher instead.
+stdout. A recording that names identity_backend is scored as that backend unless
+--identity repeats it. A recording that does not name one requires --identity;
+omitting it does not score the matcher. --identity gallery uses the same
+LiveSession path as the server and exits 2 when that gallery cannot load; it
+does not score the matcher instead.
 The report names the backend, whether the head loaded, and the weight pins.
 The same pinned recording, catalog, weights, and labels produce the
 same counts; timings are omitted because they are not.
@@ -102,9 +105,9 @@ def load_index(recording_dir: str) -> List[Dict[str, Any]]:
 def recorded_identity(rows: Sequence[Dict[str, Any]]) -> Optional[str]:
     """Backend named by the recording, when every present value agrees.
 
-    FrameRecorder does not write this. A later index may, on the first line or
-    on each frame, under identity_backend or identity.backend. Silence means
-    the caller must say --identity. A mix is not a default.
+    FrameRecorder writes identity_backend on each line when the server constructed
+    it with the active backend. Older indexes may omit it, or name it once under
+    identity.backend. Silence is not a backend. A mix is not a default.
     """
     found: List[str] = []
     for row in rows:
@@ -124,13 +127,17 @@ def recorded_identity(rows: Sequence[Dict[str, Any]]) -> Optional[str]:
 
 
 def resolve_identity(requested: Optional[str], recorded: Optional[str]) -> str:
-    """CLI wins. A recording may fill the gap. Neither one stays the matcher."""
-    chosen = (requested or recorded or "matcher").strip().lower()
+    """CLI and recording must name the same backend. Silence is not matcher."""
+    chosen = (requested or recorded or "").strip().lower()
+    if not chosen:
+        raise EvalInputError(
+            "recording does not name an identity backend; pass --identity matcher or --identity gallery"
+        )
     if chosen not in ("matcher", "gallery"):
         raise EvalInputError(f"identity must be matcher or gallery, not {requested or recorded}")
     if recorded and requested and recorded != chosen:
         raise EvalInputError(
-            f"recording identity is {recorded}; pass --identity {recorded} or omit --identity"
+            f"recording identity is {recorded}; pass --identity {recorded}"
         )
     return chosen
 
@@ -219,7 +226,7 @@ def catalog_fingerprint(stock_dir: str, stock_csv: str = "stock.csv") -> Dict[st
 def model_fingerprint(
     weights_path: str,
     *,
-    identity: str = "matcher",
+    identity: str,
     gallery: Any = None,
 ) -> Dict[str, Any]:
     """YOLO pin plus, for gallery, the backbone and head that actually loaded."""
@@ -545,7 +552,7 @@ def score_recording(
     labels_path: Optional[str] = None,
     iou_min: float = 0.5,
     live_frames: Optional[Sequence[Dict[str, Any]]] = None,
-    identity: str = "matcher",
+    identity: str,
     gallery: Any = None,
 ) -> Dict[str, Any]:
     """Score one session directory. live_frames, when given, are the replayed results.
@@ -564,7 +571,9 @@ def score_recording(
     source = list(live_frames) if live_frames is not None else rows
     tracks = [track for frame in source for track in (frame.get("tracks") or [])]
     decisions = score_tracks(tracks)
-    backend = (identity or "matcher").strip().lower()
+    backend = identity.strip().lower()
+    if backend not in ("matcher", "gallery"):
+        raise EvalInputError(f"identity must be matcher or gallery, not {identity}")
     report = {
         "recording": recording_dir,
         "source": "live_replay" if live_frames is not None else "recorded_index",
@@ -678,7 +687,7 @@ def format_report(report: Dict[str, Any]) -> str:
     head_note = f" head {head}" if head else ""
     lines = [
         f"source: {report['source']}",
-        f"identity: {identity.get('backend', 'matcher')}{head_note}",
+        f"identity: {identity.get('backend')}{head_note}",
         (
             f"decisions: accepted {decisions['accepted']}  ambiguous {decisions['ambiguous']}  "
             f"rejected {decisions['rejected']}  unknown {decisions['unknown']}  "
@@ -782,7 +791,7 @@ def main(argv: Optional[Sequence[str]] = None) -> int:
         "--identity",
         default=None,
         choices=["matcher", "gallery"],
-        help="which mechanism names products (default: recorded backend, else matcher)",
+        help="which mechanism names products (required when the recording omits identity_backend)",
     )
     parser.add_argument("--device", default=None, help="cpu | mps | cuda | cuda:0")
     args = parser.parse_args(argv)

recognization/code/finetune_identity.py

@@ -132,12 +132,14 @@ def enrollment_report_text(report: dict) -> str:
 
 
 def build_model(image_size: int = 518):
-    import timm
     import torch
 
+    from gallery import load_dinov2
+
     device = "mps" if torch.backends.mps.is_available() else "cpu"
-    model = timm.create_model("vit_small_patch14_dinov2.lvd142m", pretrained=True,
-                              num_classes=0, img_size=image_size, dynamic_img_size=True)
+    # Same hash-pinned local file as inference. pretrained=True would fetch a
+    # different copy even with YOLO_OFFLINE=1.
+    model = load_dinov2(image_size)
     return model.eval().to(device), device
 
 

recognization/code/live.py

@@ -189,8 +189,9 @@ class LiveSession:
         self._catalog_by_sku: Dict[str, Dict[str, Any]] = {}
         self._catalog_version = ""
         self._tags = None
-        # Shelf position of this frame's origin. Updated from a content shift
-        # before the session lock; tag cadence keys rows by it. Not a shelf map.
+        # Shelf position of this frame's origin. Register and origin update
+        # share one gate so overlapping workers cannot reorder either.
+        self._shift_gate = threading.Lock()
         self._frame_origin = [0.0, 0.0]
         self._shift_map = None
         self._compliance_tracker = None
@@ -337,22 +338,23 @@ class LiveSession:
         }
 
     def reset(self) -> None:
-        with self._lock:
-            self._generation += 1
-            self.tracker.reset()
-            self._tracks.clear()
-            self._last_seen_seq.clear()
-            self._relabel_count.clear()
-            self._restart_count.clear()
-            self._slot_history.clear()
-            self._label_memory.clear()
-            self._slots = []
-            self._latest_image = None
-            self._latest_frame_id = None
-            self._frame_origin = [0.0, 0.0]
-            self._shift_map = None
-            # ponytail: do not touch _waiter here. asyncio.Event is loop-affine;
-            # take_pending_async clears a stale wake on its own loop.
+        with self._shift_gate:
+            with self._lock:
+                self._generation += 1
+                self.tracker.reset()
+                self._tracks.clear()
+                self._last_seen_seq.clear()
+                self._relabel_count.clear()
+                self._restart_count.clear()
+                self._slot_history.clear()
+                self._label_memory.clear()
+                self._slots = []
+                self._latest_image = None
+                self._latest_frame_id = None
+                self._frame_origin = [0.0, 0.0]
+                self._shift_map = None
+                # ponytail: do not touch _waiter here. asyncio.Event is loop-affine;
+                # take_pending_async clears a stale wake on its own loop.
 
     # ------------------------------------------------------------------ frame path
     def process_frame(self, frame_id: int, payload: bytes) -> Dict[str, Any]:
@@ -367,7 +369,6 @@ class LiveSession:
             empty_enabled = self.settings.empty_slots
             generation = self._generation
 
-
         shelf_offset = self._note_frame_shift(image)
 
         boxes = self.engine.detect(image, conf=conf)
@@ -495,7 +496,7 @@ class LiveSession:
             summary = self._summary_locked()
             stats = self._stats_locked()
             track_payload = [self._track_payload(state) for state in self._tracks.values()]
-            tag_summary = self._apply_tags(seq, track_payload, tag_bands)
+            tag_summary = self._apply_tags(seq, track_payload, tag_bands, shelf_offset, image.shape)
             if tag_summary is not None:
                 summary["price_tags"] = tag_summary["tags"]
                 summary["compliance"] = tag_summary["compliance"]
@@ -996,16 +997,19 @@ class LiveSession:
     def _note_frame_shift(self, image: Any) -> Tuple[float, float]:
         """Shelf position of this frame's origin. Phase correlation only.
 
-        Runs before the session lock. A failed registration keeps the last origin
-        rather than inventing a jump. The content shift moves the origin the
-        other way, same sign as ShelfMap.
+        Register and origin update share one per-session gate, so overlapping
+        workers cannot reorder the previous-frame baseline or publish a later
+        origin first. Correlation stays outside the session lock and outside
+        OCR. A failed registration keeps the last origin rather than inventing
+        a jump. The content shift moves the origin the other way, same sign as
+        ShelfMap.
         """
-        if self._shift_map is None:
-            from shelf_map import ShelfMap
+        with self._shift_gate:
+            if self._shift_map is None:
+                from shelf_map import ShelfMap
 
-            self._shift_map = ShelfMap()
-        shift = self._shift_map.register(image)
-        with self._lock:
+                self._shift_map = ShelfMap()
+            shift = self._shift_map.register(image)
             if shift is not None:
                 self._frame_origin[0] -= float(shift[0])
                 self._frame_origin[1] -= float(shift[1])
@@ -1034,19 +1038,27 @@ class LiveSession:
             LOGGER.warning("tag pass failed: %s", error)
 
     def _apply_tags(
-        self, seq: int, track_payloads: List[Dict[str, Any]], bands: Sequence[Any]
+        self,
+        seq: int,
+        track_payloads: List[Dict[str, Any]],
+        bands: Sequence[Any],
+        shelf_offset: Sequence[float],
+        image_shape: Sequence[int],
     ) -> Optional[Dict[str, Any]]:
         """Attach tag + reconciliation to each track and compute compliance.
 
         Must be called with the session lock held (it reads the tag cache under its
-        own lock, which is never taken the other way round).
+        own lock, which is never taken the other way round). The origin and shape
+        are this frame's, not whatever `plan` last stored.
         """
         service = self._tags
         if service is None or not self.settings.tags:
             return None
         from reconcile import ComplianceTracker, evaluate_compliance, reconcile, summarize_findings
 
-        tag_by_track, all_tags, counts = service.attach(track_payloads)
+        tag_by_track, all_tags, counts = service.attach(
+            track_payloads, shelf_offset=shelf_offset, image_shape=image_shape
+        )
         reconciliations: Dict[int, Dict[str, Any]] = {}
         for payload in track_payloads:
             track_id = int(payload["track_id"])

recognization/code/record.py

@@ -15,7 +15,8 @@ Layout:
 
     DIR/<session>/000123.jpg          the frame, exactly as the client sent it
     DIR/<session>/index.jsonl         one line per frame:
-        {"frame": 123, "path": "...", "at": 1699..., "summary": {...},
+        {"frame": 123, "path": "...", "at": 1699..., "identity_backend": "gallery",
+         "summary": {...},
          "tracks": [{"box": [...], "identity": {...}, "tag": {...}}]}
 
 Frames are written at most every `every` processed frames (default 3) so a long
@@ -35,10 +36,17 @@ from typing import Any, Dict, List, Optional
 class FrameRecorder:
     """Write frames and their results to disk, for labelling later."""
 
-    def __init__(self, root: str, every: int = 3, max_frames_per_session: int = 4000) -> None:
+    def __init__(
+        self,
+        root: str,
+        every: int = 3,
+        max_frames_per_session: int = 4000,
+        identity_backend: Optional[str] = None,
+    ) -> None:
         self.root = root
         self.every = max(1, int(every))
         self.max_frames_per_session = int(max_frames_per_session)
+        self.identity_backend = (identity_backend or "").strip().lower() or None
         self._lock = threading.Lock()
         self._counts: Dict[str, int] = {}
         self._written: Dict[str, int] = {}
@@ -86,6 +94,8 @@ class FrameRecorder:
             "summary": result.get("summary"),
             "tracks": tracks,
         }
+        if self.identity_backend:
+            line["identity_backend"] = self.identity_backend
         try:
             with open(os.path.join(directory, "index.jsonl"), "a", encoding="utf-8") as handle:
                 handle.write(json.dumps(line, ensure_ascii=False) + "\n")

recognization/code/tags.py

@@ -788,16 +788,20 @@ class TagService:
         Records live in shelf coordinates. They are projected into this frame
         before association, and a tag that does not overlap the frame is left
         out, so an old tag cannot bind a new facing that reused its screen box.
-        Returned boxes stay in screen pixels. `shelf_offset` defaults to the
-        origin of the latest `plan` (a still, or a caller that has not panned).
+        Returned boxes stay in screen pixels. A caller that omits `shelf_offset`
+        is a still: screen coordinates, never another frame's stored origin.
 
         Returns (tag_by_track_id, all_tags, state_counts). A track whose tag was
         never read gets state 'unreadable' when a tag was found near it, and
         'absent' when nothing tag-shaped was seen there.
         """
         with self._lock:
-            origin = _origin_of(shelf_offset if shelf_offset is not None else self._view_origin)
-            shape = image_shape if image_shape is not None else self._view_shape
+            origin = _origin_of(shelf_offset)
+            shape = (
+                (int(image_shape[0]), int(image_shape[1]))
+                if image_shape is not None
+                else self._view_shape
+            )
             visible = _visible_records(list(self._tags.values()), origin, shape)
 
         by_track: Dict[int, Dict[str, Any]] = {}

recognization/code/test_engine.py

@@ -274,6 +274,33 @@ class CatalogTests(unittest.TestCase):
         self.assertEqual(product["number_out"], 3)
         self.assertIsNone(engine.update_stock("nope", -1))
 
+
+class GalleryLoadTests(unittest.TestCase):
+    def test_a_failed_gallery_load_returns_none_and_is_not_retried(self):
+        engine, _, _, _ = make_engine(identity_backend="gallery", stock_dir="does-not-exist")
+
+        class BrokenGallery:
+            def __init__(self, *args, **kwargs):
+                pass
+
+            def load(self):
+                raise FileNotFoundError("DINOv2 weights not found")
+
+        import gallery as gallery_module
+
+        original = gallery_module.ProductGallery
+        gallery_module.ProductGallery = BrokenGallery
+        try:
+            self.assertIsNone(engine.identity_gallery())
+            self.assertIsNone(engine.identity_gallery())
+        finally:
+            gallery_module.ProductGallery = original
+            engine.shutdown()
+        self.assertIs(engine._gallery, False)
+
+
+
+class CatalogImageTests(unittest.TestCase):
     def test_catalog_image_path_is_resolved_through_inventory(self):
         engine, _, _, _ = make_engine()
         self.assertIsNone(engine.catalog_image_path("../../etc/passwd"))

recognization/code/test_eval_live.py

@@ -23,6 +23,8 @@ from eval_live import (  # noqa: E402
     labelled_top1,
     leaked_labels,
     load_index,
+    recorded_identity,
+    resolve_identity,
     score_recording,
     score_tracks,
     track_churn,
@@ -176,7 +178,7 @@ class LeakageTests(unittest.TestCase):
             )
         with open(os.path.join(recording, "labels.json"), "w", encoding="utf-8") as handle:
             json.dump({"labels": [{"frame": 1, "box": [0, 0, 12, 12], "sku_id": "sku_a"}]}, handle)
-        report = score_recording(recording, stock_dir=stock)
+        report = score_recording(recording, stock_dir=stock, identity="matcher")
         self.assertEqual(report["leakage"]["excluded"], 1)
         self.assertEqual(report["labelled_top1"]["labels"], 0)
         self.assertEqual(report["labelled_top1"]["top1"], None)
@@ -220,8 +222,8 @@ class CliTests(unittest.TestCase):
             handle.write(json.dumps(line) + "\n")
         with open(os.path.join(recording, "labels.json"), "w", encoding="utf-8") as handle:
             json.dump({"labels": [{"frame": 4, "box": [0, 0, 8, 8], "sku_id": "sku_a"}]}, handle)
-        first = score_recording(recording, stock_dir=stock)
-        second = score_recording(recording, stock_dir=stock)
+        first = score_recording(recording, stock_dir=stock, identity="matcher")
+        second = score_recording(recording, stock_dir=stock, identity="matcher")
         self.assertEqual(first, second)
         self.assertEqual(first["decisions"]["accepted"], 1)
         self.assertEqual(first["decisions"]["ambiguous"], 1)
@@ -273,5 +275,26 @@ class IdentityContractTests(unittest.TestCase):
         self.assertFalse(os.path.exists(report_path))
         self.assertNotIn("accepted_rate", completed.stdout)
 
+    def test_a_silent_recording_does_not_become_the_matcher(self):
+        with self.assertRaises(EvalInputError):
+            resolve_identity(None, None)
+        self.assertEqual(resolve_identity("matcher", None), "matcher")
+        self.assertEqual(resolve_identity(None, "gallery"), "gallery")
+        with self.assertRaises(EvalInputError):
+            resolve_identity("matcher", "gallery")
+
+    def test_recorded_backend_is_read_and_a_mix_is_refused(self):
+        self.assertEqual(recorded_identity([{"identity_backend": "Gallery"}]), "gallery")
+        self.assertIsNone(recorded_identity([{"frame": 1}]))
+        self.assertEqual(
+            recorded_identity([{"identity": {"backend": "matcher", "sku_id": "sku_a"}}]),
+            "matcher",
+        )
+        with self.assertRaises(EvalInputError):
+            recorded_identity([
+                {"identity_backend": "gallery"},
+                {"identity_backend": "matcher"},
+            ])
+
 if __name__ == "__main__":
     unittest.main()

recognization/code/test_gallery.py

@@ -225,5 +225,43 @@ class HeadPinTests(unittest.TestCase):
             self.assertEqual(gallery._head["class_ids"], subset)
             self.assertIsNone(head_class_set(["sku_a", "sku_a", ""]))
 
+
+class TrainerPinTests(unittest.TestCase):
+    def test_head_training_loads_the_pinned_file_and_keeps_preprocess(self):
+        import finetune_identity
+        import gallery as gallery_module
+
+        seen = []
+
+        class FakeBackbone:
+            def eval(self):
+                return self
+
+            def to(self, device):
+                return self
+
+        sentinel = FakeBackbone()
+        def fake_load(image_size):
+            seen.append(image_size)
+            return sentinel
+
+
+        original = finetune_identity.load_dinov2
+        finetune_identity.load_dinov2 = fake_load
+        try:
+            model, device = finetune_identity.build_model(224)
+        finally:
+            finetune_identity.load_dinov2 = original
+        self.assertIs(model, sentinel)
+        self.assertIn(device, ("cpu", "mps"))
+        self.assertEqual(seen, [224])
+
+        image = np.zeros((10, 20, 3), dtype=np.uint8)
+        self.assertEqual(finetune_identity.prepare(image, 32).shape, gallery_module._prepare(image, 32).shape)
+        self.assertTrue(np.allclose(
+            finetune_identity.prepare(image, 32),
+            gallery_module._prepare(image, 32),
+        ))
+
 if __name__ == "__main__":
     unittest.main()

recognization/code/test_live.py

@@ -307,6 +307,80 @@ class TagPassIntegrationTests(unittest.TestCase):
         )
         engine.shutdown()
 
+    def test_overlapping_frames_keep_their_own_origin(self):
+        """A later worker must not become the previous-frame baseline or steal origin."""
+        session, engine, _ = self.make_tagged_session()
+
+        class Ordered:
+            def __init__(self):
+                self.inside = 0
+                self.max_inside = 0
+                self.calls = 0
+                self._inner = threading.Lock()
+
+            def register(self, image):
+                with self._inner:
+                    self.inside += 1
+                    self.max_inside = max(self.max_inside, self.inside)
+                    self.calls += 1
+                    call = self.calls
+                time.sleep(0.03)
+                with self._inner:
+                    self.inside -= 1
+                return (0.0, -400.0)
+
+        session._shift_map = Ordered()
+        errors = []
+
+        def run(frame_id):
+            try:
+                session.process_frame(frame_id, frame_bytes())
+            except Exception as exc:
+                errors.append(exc)
+
+        early = threading.Thread(target=run, args=(1,))
+        late = threading.Thread(target=run, args=(2,))
+        early.start()
+        late.start()
+        early.join(2.0)
+        late.join(2.0)
+        self.assertEqual(errors, [])
+        self.assertFalse(early.is_alive() or late.is_alive())
+        self.assertEqual(session._shift_map.calls, 2)
+        self.assertEqual(session._shift_map.max_inside, 1)
+        self.assertEqual(session._frame_origin[1], 800.0)
+        engine.shutdown()
+
+    def test_cached_tag_projects_at_this_frame_origin(self):
+        """A panned frame must not attach a tag that only overlaps the previous origin."""
+        from tags import TagRecord
+
+        session, engine, _ = self.make_tagged_session()
+        record = TagRecord(
+            tag_id="cached", box=[150, 700, 190, 760], band_index=0, name="cached", price=150,
+            flags=["tax_consistent"],
+        )
+        record.settled = True
+        session._tags._tags[record.tag_id] = record
+
+        class Panned:
+            def __init__(self):
+                self.calls = 0
+
+            def register(self, image):
+                self.calls += 1
+                return (0.0, -400.0) if self.calls == 2 else (0.0, 0.0)
+
+        session._shift_map = Panned()
+        first = session.process_frame(1, frame_bytes())
+        self.assertFalse(any(track.get("price_tag", {}).get("id") for track in first["tracks"]))
+        second = session.process_frame(2, frame_bytes())
+        bound = [track["price_tag"] for track in second["tracks"] if track.get("price_tag", {}).get("id")]
+        self.assertTrue(bound)
+        self.assertGreaterEqual(bound[0]["box"][1], 200)
+        self.assertLess(bound[0]["box"][1], 480)
+        engine.shutdown()
+
 
 class FramePipelineTests(unittest.TestCase):
     def test_latest_frame_wins_and_drops_are_counted(self):

recognization/code/test_ocr_tags.py

@@ -385,6 +385,27 @@ class TagServiceTests(unittest.TestCase):
         self.assertEqual(by_track[9]["state"], TAG_STATE_ABSENT)
         self.assertIsNone(by_track[9]["id"])
 
+    def test_attach_does_not_use_the_latest_plan_origin(self):
+        """A later plan must not project this attach through its stored origin."""
+        service, _ = self.make(pass_interval_frames=0, settle_observations=1, settle_passes=1)
+        boxes = [box(100, 100, 200, 300, 0), box(220, 100, 320, 300, 1), box(340, 100, 440, 300, 2)]
+        array = image()
+        first = service.plan(boxes, array.shape, shelf_offset=(0.0, 0.0))
+        service.run_pass(array, first[0])
+        service.plan(boxes, array.shape, shelf_offset=(0.0, 400.0))
+        by_track, tags, _ = service.attach(
+            [{"track_id": 0, "box": [100, 100, 200, 300]}],
+            image_shape=array.shape,
+        )
+        self.assertTrue(tags, "omitting shelf_offset is screen coordinates, not the panned plan origin")
+        self.assertEqual(by_track[0]["state"], TAG_STATE_READABLE)
+        by_track, tags, _ = service.attach(
+            [{"track_id": 0, "box": [100, 100, 200, 300]}],
+            shelf_offset=(0.0, 0.0),
+            image_shape=array.shape,
+        )
+        self.assertEqual(by_track[0]["state"], TAG_STATE_READABLE)
+
 
 class StorePolicyTests(unittest.TestCase):
     def test_a_missing_file_uses_the_defaults(self):

recognization/code/test_record.py

@@ -45,6 +45,21 @@ class RecorderTests(unittest.TestCase):
         self.assertEqual([line["frame"] for line in lines], [0, 1, 2])
         self.assertEqual(lines[0]["tracks"][0]["tag_identity"]["sku_id"], "sku_l05")
 
+    def test_the_active_backend_is_written_on_each_line(self):
+        recorder = FrameRecorder(self.root, every=1, identity_backend="gallery")
+        recorder.write("session-d", 0, b"x", a_result("sku_l05"))
+        recorder.write("session-d", 1, b"y", a_result())
+        with open(os.path.join(self.root, "session-d", "index.jsonl"), encoding="utf-8") as handle:
+            lines = [json.loads(line) for line in handle]
+        self.assertEqual([line["identity_backend"] for line in lines], ["gallery", "gallery"])
+
+    def test_an_unspecified_backend_is_omitted(self):
+        recorder = FrameRecorder(self.root, every=1)
+        recorder.write("session-e", 0, b"x", a_result())
+        with open(os.path.join(self.root, "session-e", "index.jsonl"), encoding="utf-8") as handle:
+            line = json.loads(handle.readline())
+        self.assertNotIn("identity_backend", line)
+
     def test_sampling_keeps_the_frame_numbering_visible(self):
         recorder = FrameRecorder(self.root, every=3)
         for frame_id in range(9):

recognization/docs/DATA.md

@@ -38,6 +38,7 @@ Then run the app and scan. Each scan session becomes its own directory:
     000001.jpg
     index.jsonl       one line per recorded frame:
                       {"frame": 123, "path": "000001.jpg", "at": 1699...,
+                       "identity_backend": "gallery",
                        "summary": {...},
                        "tracks": [{"box": [...], "identity": {...},
                                    "tag_identity": {...}, "tag": {...}}]}
@@ -97,9 +98,15 @@ predictions. For the second visit, run from `recognization` with the same frozen
 SESSION="$HOME/dorayaki-data/storeA-visit2/REPLACE_WITH_SESSION"
 YOLO_OFFLINE=1 .venv/bin/python code/eval_live.py "$SESSION" \
   --stock-dir stock-lopia --labels "$SESSION/labels.json" \
-  --replay --identity gallery --device cpu
+  --replay --device cpu
 ```
 
+A recording written by the current server names `identity_backend` on each
+index line, and the evaluator uses that. `--identity gallery` (or `matcher`)
+is required when the index omits it, including recordings from before that
+field existed, and must match the recorded value when both are present. It
+does not silently score the matcher.
+
 Replace `REPLACE_WITH_SESSION` with the actual session name. The evaluator replays
 recorded JPEGs through `LiveSession`, writes `eval_live.json` in that directory,
 and reports backend, loaded head, weight/catalog fingerprints, decision counts,