fix: sync DB sessions in Celery, letterbox YuNet, dedupe detections

- Rewrite all vision tasks to use sync psycopg2 sessions instead of
  asyncpg — fixes 'another operation in progress' and event loop errors
  when Celery forks workers sharing the async connection pool
- Letterbox-pad images to exactly 640x640 for YuNet face detector
  (was crashing on non-square thumbnails)
- Deduplicate object detections per label per photo — keep highest
  confidence only to avoid photo_tags PK violation on multiple
  detections of the same class
- Add all queues (-Q default,high,low,vision) to worker command

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

backend/app/services/vision/faces.py

@@ -68,7 +68,7 @@ class YuNetSFaceProcessor(FaceProcessor):
     def process(self, image: np.ndarray) -> list[FaceDetection]:
         orig_h, orig_w = image.shape[:2]
 
-        # Scale image for YuNet (expects fixed input size)
+        # Scale + letterbox to exactly 640x640 (YuNet fixed input)
         scale = min(_YUNET_INPUT_SIZE / orig_w, _YUNET_INPUT_SIZE / orig_h)
         new_w = int(orig_w * scale)
         new_h = int(orig_h * scale)
@@ -79,12 +79,17 @@ class YuNetSFaceProcessor(FaceProcessor):
             dtype=np.uint8,
         )
 
-        # YuNet expects BGR, uint8, NHWC
-        bgr = resized[:, :, ::-1].copy()
+        # Letterbox pad to 640x640
+        canvas = np.full((_YUNET_INPUT_SIZE, _YUNET_INPUT_SIZE, 3), 128, dtype=np.uint8)
+        pad_y = (_YUNET_INPUT_SIZE - new_h) // 2
+        pad_x = (_YUNET_INPUT_SIZE - new_w) // 2
+        canvas[pad_y:pad_y + new_h, pad_x:pad_x + new_w] = resized
+
+        # YuNet expects BGR
+        bgr = canvas[:, :, ::-1].copy()
 
         # Run detection
         det_input = self._detector.get_inputs()[0]
-        # YuNet uses dynamic input — reshape
         blob = bgr.astype(np.float32)[np.newaxis]  # (1, H, W, 3)
         # Some YuNet ONNX exports expect (1, 3, H, W)
         if det_input.shape and len(det_input.shape) == 4 and det_input.shape[1] == 3:
@@ -109,11 +114,11 @@ class YuNetSFaceProcessor(FaceProcessor):
             if face_size < self._min_face_size:
                 continue
 
-            # Rescale to original image coords
-            x1 = x / scale
-            y1 = y / scale
-            x2 = (x + w) / scale
-            y2 = (y + h) / scale
+            # Remove letterbox padding and rescale to original image coords
+            x1 = (x - pad_x) / scale
+            y1 = (y - pad_y) / scale
+            x2 = (x + w - pad_x) / scale
+            y2 = (y + h - pad_y) / scale
 
             bbox = [
                 max(0, x1 / orig_w),
@@ -124,7 +129,9 @@ class YuNetSFaceProcessor(FaceProcessor):
 
             # Extract landmarks (5 points) for alignment
             if len(det) >= 15:
-                landmarks = det[5:15].reshape(5, 2) / scale
+                landmarks = det[5:15].reshape(5, 2)
+                landmarks[:, 0] = (landmarks[:, 0] - pad_x) / scale
+                landmarks[:, 1] = (landmarks[:, 1] - pad_y) / scale
             else:
                 # Fallback: no landmarks, skip recognition
                 continue