Improve OCR preprocessing and amount extraction robustness

Image preprocessing (receipt_parser.py):
- Add ImageOps.exif_transpose() — fixes portrait photos stored with EXIF
  rotation metadata (most phone photos); without this Tesseract reads a
  rotated image and produces garbage
- Upscale images < 600px wide for better character recognition
- Raise binarization threshold 140→160 for faint thermal-print receipts
- Try PSM 6 (single text block) before PSM 4, PSM 11 as fallbacks;
  PSM 6 is better suited to single-column receipt layout

Amount extraction (expenses_agent.py):
- Add Pass 2 bottom-of-receipt line scan when labeled Total: regex fails;
  reads lines bottom-to-top in the last 50% of text, skipping change/tip
  lines — handles 'T0TAL' OCR misread and amount-on-next-line layout
- Add _SKIP_LINE_RE and _ANY_DOLLAR_RE module-level patterns
- 8 new tests covering garbled total, change-skip, USD suffix, etc.

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

agent_service/tools/receipt_parser.py

@@ -85,34 +85,58 @@ def _ocr_image(data: bytes, filename: str) -> str:
 
 
 def _ocr_image_tesseract(data: bytes, filename: str) -> str:
-    """Tesseract-based OCR pipeline (fallback)."""
+    """Tesseract-based OCR pipeline with phone-photo preprocessing."""
     try:
         from PIL import Image, ImageFilter, ImageOps
         import pytesseract
         img = Image.open(io.BytesIO(data))
 
-        # Resize very large images — tesseract is slower and less accurate at
-        # phone-camera resolution; 1800px wide is plenty for receipt text.
+        # ── Step 1: EXIF rotation correction ─────────────────────────────────
+        # Phone photos are stored with EXIF orientation metadata but the pixel
+        # data is not actually rotated.  Without this fix Tesseract reads a
+        # portrait receipt as a landscape image and produces garbage.
+        try:
+            img = ImageOps.exif_transpose(img)
+        except Exception:
+            pass  # exif_transpose requires Pillow >= 6.0
+
+        # ── Step 2: Resize to working width (1800px) ──────────────────────────
         max_w = 1800
         if img.width > max_w:
             scale = max_w / img.width
             img = img.resize((max_w, int(img.height * scale)), Image.LANCZOS)
+        # Upscale very small images — Tesseract accuracy drops below ~600px
+        elif img.width < 600:
+            scale = 600 / img.width
+            img = img.resize((600, int(img.height * scale)), Image.LANCZOS)
 
-        # Grayscale + adaptive binarisation + sharpen
+        # ── Step 3: Grayscale + contrast ─────────────────────────────────────
         img = ImageOps.grayscale(img)
         img = ImageOps.autocontrast(img)
-        img = img.point(lambda x: 0 if x < 140 else 255)
+
+        # ── Step 4: Sharpen then binarize ─────────────────────────────────────
+        # Sharpen first so edges are crisp before thresholding.
+        # Threshold 160 (was 140) — gentler for faint thermal-print receipts
+        # where light gray text would be wiped out by the stricter threshold.
         img = img.filter(ImageFilter.SHARPEN)
+        img = img.point(lambda x: 0 if x < 160 else 255)
 
-        # psm 1 = automatic page segmentation + OSD (handles rotated receipts).
-        # Fall back to psm 6 if OSD data is missing.
-        try:
-            text = pytesseract.image_to_string(img, config='--oem 3 --psm 1').strip()
-        except Exception:
-            text = pytesseract.image_to_string(img, config='--oem 3 --psm 6').strip()
+        # ── Step 5: OCR — try PSM modes best-suited for receipt layout ────────
+        # PSM 6 = single uniform text block (best for single-column receipts)
+        # PSM 4 = single column, variable text sizes (wider fallback)
+        # PSM 11 = sparse text — last resort for badly segmented images
+        for psm in (6, 4, 11):
+            try:
+                text = pytesseract.image_to_string(
+                    img, config=f'--oem 3 --psm {psm}').strip()
+                if len(text) >= 20:
+                    logger.debug('Tesseract OCR %s: psm=%d %d chars', filename, psm, len(text))
+                    return text
+            except Exception:
+                pass
 
-        logger.debug('Tesseract OCR %s: %d chars', filename, len(text))
-        return text
+        logger.warning('Tesseract OCR %s: all PSM modes returned < 20 chars', filename)
+        return ''
     except ImportError:
         logger.warning('pytesseract/Pillow not installed — OCR unavailable for %s', filename)
         return f'[Image: {filename} — install pytesseract+Pillow for OCR]'