Add deduplication: keep image closest to camera within 2s window

Within a burst sequence from the same station, the image where the athlete
is physically closest to the camera gives the most accurate passage timestamp.
Proximity is measured by bib bounding box area (larger = closer).

When a duplicate is detected:
- New image closer: update timestamp + image path, delete old image
- Existing image closer: discard new image

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

backend/ocr.py

@@ -29,6 +29,7 @@ class OcrResult:
     digits: Optional[str]       # Gjenkjente sifre, f.eks. "42", None hvis ingen
     confidence: float           # 0.0–1.0
     partial: bool               # True hvis nummeret trolig er delvis skjult
+    proximity_score: float = 0.0  # Areal av detektert bib-boks (px²) — større = nærmere kamera
     raw_texts: list[str] = field(default_factory=list)  # Alle OCR-treff for debug
 
 
@@ -56,28 +57,34 @@ def _preprocess(image_path: Path) -> np.ndarray:
     return enhanced
 
 
-def _extract_bib_number(texts: list[tuple]) -> tuple[Optional[str], float, bool]:
+def _bbox_area(bbox) -> float:
+    """Beregn areal av EasyOCR bounding box [[x1,y1],[x2,y2],[x3,y3],[x4,y4]]."""
+    xs = [p[0] for p in bbox]
+    ys = [p[1] for p in bbox]
+    return (max(xs) - min(xs)) * (max(ys) - min(ys))
+
+
+def _extract_bib_number(texts: list[tuple]) -> tuple[Optional[str], float, bool, float]:
     """
     Finn beste siffersekvens blant OCR-treff.
-    Returnerer (sifre, konfidens, partial).
+    Returnerer (sifre, konfidens, partial, proximity_score).
+    proximity_score = areal av bounding box i piksler² (større = nærmere kamera).
     """
     candidates = []
-    for (_, text, conf) in texts:
-        # Behold kun sifre
+    for (bbox, text, conf) in texts:
         digits = re.sub(r"[^0-9]", "", text)
         if digits:
-            candidates.append((digits, float(conf)))
+            candidates.append((digits, float(conf), _bbox_area(bbox)))
 
     if not candidates:
-        return None, 0.0, False
+        return None, 0.0, False, 0.0
 
     # Velg kandidat med høyest konfidens
-    best_digits, best_conf = max(candidates, key=lambda x: x[1])
+    best_digits, best_conf, best_area = max(candidates, key=lambda x: x[1])
 
-    # Heuristikk: 1–2 sifre kan tyde på delvis synlig nummer
     partial = len(best_digits) < 2
 
-    return best_digits, best_conf, partial
+    return best_digits, best_conf, partial, best_area
 
 
 def read_bib(image_path: Path) -> OcrResult:
@@ -91,12 +98,13 @@ def read_bib(image_path: Path) -> OcrResult:
         results = reader.readtext(processed, detail=1, paragraph=False)
 
         raw_texts = [text for (_, text, _) in results]
-        digits, confidence, partial = _extract_bib_number(results)
+        digits, confidence, partial, proximity_score = _extract_bib_number(results)
 
         return OcrResult(
             digits=digits,
             confidence=confidence,
             partial=partial,
+            proximity_score=proximity_score,
             raw_texts=raw_texts,
         )
     except Exception as e: