expenses_agent: fix OCR '$→8' misread inflating receipt totals

Add _fix_ocr_dollar_as_8() which strips a spurious leading '8' when it
sits at a word boundary before a non-zero digit + 1–3 more digits + .dd
(covers $10–$9999).  Applied at the top of _extract_amount_from_text so
both the labeled-total pass and the max-scan pass benefit.

  845.00  → 45.00   ($45 misread as 845)
  885.00  → 85.00   ($85 misread as 885)
  8150.00 → 150.00  ($150 misread as 8150)
  85.00   → 85.00   UNCHANGED (real $85 correctly read)
  8.50    → 8.50    UNCHANGED (real $8.50 correctly read)

12 new tests covering fix cases, non-fix cases, and end-to-end extraction
(110 tests total, all passing).

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

agent_service/agents/expenses_agent.py

@@ -29,6 +29,27 @@ _TOTAL_RE = re.compile(
     re.IGNORECASE,
 )
 
+# OCR artefact: the '$' glyph is often misclassified as '8', turning
+# 'Total: $45.00' into 'Total: 845.00'.  We strip the spurious leading '8'
+# when it sits at a word boundary and is followed by a non-zero digit then
+# 1-3 more digits + two decimal places.  This covers the $10–$9999 range.
+#
+#   845.00  → 45.00   (was $45, OCR gave 845)
+#   885.00  → 85.00   (was $85, OCR gave 885)
+#   8150.00 → 150.00  (was $150, OCR gave 8150)
+#   85.00   → 85.00   UNCHANGED — real $85 correctly read
+#   8.50    → 8.50    UNCHANGED — real $8.50 correctly read
+#   12845.00→ 12845.00 UNCHANGED — digit before the 8 blocks lookbehind
+# Edge case: a real $8xx amount correctly read (e.g. 840.00) may be reduced
+# to $40; this is rare compared to the misread and obvious on human review.
+_OCR_DOLLAR_MISREAD_RE = re.compile(r'(?<!\d)8([1-9]\d{1,3}\.\d{2})\b')
+
+
+def _fix_ocr_dollar_as_8(text: str) -> str:
+    """Strip a spurious leading '8' that is an OCR misread of '$'."""
+    return _OCR_DOLLAR_MISREAD_RE.sub(r'\1', text)
+
+
 # Lines that should never be treated as the total — change given back,
 # tip added after the fact, etc.  Card-brand lines like "VISA USD$ 36.78"
 # are intentionally NOT listed here: the amount on those lines IS the charge.
@@ -109,6 +130,9 @@ def _extract_amount_from_text(text: str) -> float:
     if not text:
         return 0.0
 
+    # Normalise '$→8' OCR misread before any pattern matching.
+    text = _fix_ocr_dollar_as_8(text)
+
     # Pass 1: explicit label match — return the LARGEST labeled amount.
     # Using max() rather than the last positional match handles the common
     # OCR artefact where "Total\n$2.80" (garbled "Total Taxes") appears

tests/test_expenses_agent.py

@@ -428,7 +428,7 @@ async def test_act_no_employee_returns_empty_and_escalates():
 
 from agent_service.agents.expenses_agent import (
     _extract_amount_from_text, _extract_date_from_text, _is_likely_bank_statement,
-    _MONTH_MAP, _get_vision_mode,
+    _fix_ocr_dollar_as_8, _MONTH_MAP, _get_vision_mode,
 )
 
 
@@ -537,6 +537,54 @@ class TestExtractAmount:
         assert _extract_amount_from_text(text) == 150.00
 
 
+class TestFixOcrDollarAs8:
+    """_fix_ocr_dollar_as_8 — strip spurious leading '8' OCR misread of '$'."""
+
+    def test_two_digit_amount_fixed(self):
+        # $45 misread as 845
+        assert _fix_ocr_dollar_as_8('Total: 845.00') == 'Total: 45.00'
+
+    def test_three_digit_amount_fixed(self):
+        # $150 misread as 8150
+        assert _fix_ocr_dollar_as_8('Amount: 8150.00') == 'Amount: 150.00'
+
+    def test_misread_dollar_85_fixed(self):
+        # $85 misread as 885 — result should be 85.00
+        assert _fix_ocr_dollar_as_8('Total: 885.00') == 'Total: 85.00'
+
+    def test_real_85_not_touched(self):
+        # Real $85 correctly read as 85.00 — 8 followed by only 1 digit → no match
+        assert _fix_ocr_dollar_as_8('Total: 85.00') == 'Total: 85.00'
+
+    def test_real_8_50_not_touched(self):
+        # Single-digit dollar amount, correctly read — unchanged
+        assert _fix_ocr_dollar_as_8('Price: 8.50') == 'Price: 8.50'
+
+    def test_explicit_dollar_sign_not_touched(self):
+        # Already has '$' — nothing to fix
+        assert _fix_ocr_dollar_as_8('Total: $45.00') == 'Total: $45.00'
+
+    def test_digit_before_8_not_touched(self):
+        # 8 is mid-number (preceded by digit) — lookbehind blocks it
+        assert _fix_ocr_dollar_as_8('Amount: 12845.00') == 'Amount: 12845.00'
+
+    def test_zero_leading_digit_not_touched(self):
+        # 8 followed by 0XX — could be real $8000; left alone
+        assert _fix_ocr_dollar_as_8('Total: 8045.00') == 'Total: 8045.00'
+
+    def test_extract_amount_applies_fix(self):
+        # Integration: extraction uses the fix internally
+        assert _extract_amount_from_text('TOTAL 845.00') == 45.00
+
+    def test_extract_amount_labeled_total_fixed(self):
+        assert _extract_amount_from_text('Grand Total: 8150.00') == 150.00
+
+    def test_extract_amount_pass2_scan_fixed(self):
+        # No 'Total' label — Pass 2 max scan must also see the fixed amount
+        text = 'BURGER  12.99\nFRIES    4.50\n845.00'
+        assert _extract_amount_from_text(text) == 45.00
+
+
 class TestBankStatementDetection:
     def _stmt(self, n: int) -> str:
         """Generate fake bank statement with n transaction lines."""