fix: improve OCR accuracy for rotated/sideways receipt photos

- Dockerfile: add tesseract-ocr-osd for orientation detection data
- receipt_parser: resize large phone photos to 1800px, convert to
  grayscale, sharpen before OCR; use psm 1 (auto + OSD) so rotated
  receipts are correctly oriented before text extraction
- expenses_agent: tighten amount extraction prompt to pick the FINAL
  total, not subtotal or tax line, reducing misreads like 42.90->409.00

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

agent_service/agents/expenses_agent.py

@@ -220,11 +220,14 @@ class ExpensesAgent(BaseAgent):
             prompt = (
                 'Extract expense details from the following receipt text. '
                 'Return ONLY valid JSON with these keys:\n'
-                '"vendor" (string, merchant name),\n'
-                '"amount" (number, the total amount charged — look for "Total", "Amount Due", "Grand Total"),\n'
-                f'"date" (string YYYY-MM-DD, use {date_hint or today} if not found),\n'
+                '"vendor" (string, merchant or restaurant name),\n'
+                '"amount" (number — the FINAL total the customer paid; '
+                'this is labeled "Total", "Amount Due", "Grand Total", or the last dollar figure; '
+                'do NOT use subtotal, tax, or tip separately; '
+                'if multiple totals appear pick the largest one labeled as the final total),\n'
+                f'"date" (string YYYY-MM-DD, use {date_hint or today} if not found in text),\n'
                 f'"product_name" (string, pick the best match from [{product_list}] or empty string).\n\n'
-                f'Receipt text (first 2000 chars):\n{text[:2000]}\n\nJSON only:'
+                f'Receipt text:\n{text[:2000]}\n\nJSON only:'
             )
         try:
             resp = await self._llm.submit(