Fix receipt parsing quality and approval endpoint

Receipt quality: replace LLM amount/date extraction with regex.
LLM was hallucinating 2021/2022 dates and returning '198.40 USD' strings.
Amounts now use deterministic regex (Total:/Grand Total:/Amount Due:).
Dates: filename timestamp > OCR regex > today (no LLM date guessing).
LLM only asked for vendor name + product category.

Approval: fix GET /approval/pending 500 by using correct column
name 'started_at' instead of 'created_at' (which does not exist).

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

tests/test_expenses_agent.py

@@ -423,15 +423,73 @@ async def test_act_no_employee_returns_empty_and_escalates():
 
 
 # ---------------------------------------------------------------------------
-# _parse_receipt_text — LLM extraction path
+# _extract_amount_from_text / _extract_date_from_text — regex helpers
+# ---------------------------------------------------------------------------
+
+from agent_service.agents.expenses_agent import (
+    _extract_amount_from_text, _extract_date_from_text,
+)
+
+
+class TestExtractAmount:
+    def test_simple_total(self):
+        assert _extract_amount_from_text('Acme\nTotal: $9.99') == 9.99
+
+    def test_grand_total(self):
+        assert _extract_amount_from_text('Subtotal: $20.00\nGrand Total: $22.46') == 22.46
+
+    def test_amount_due(self):
+        assert _extract_amount_from_text('Amount Due: 198.40') == 198.40
+
+    def test_no_dollar_sign(self):
+        assert _extract_amount_from_text('TOTAL 15.75') == 15.75
+
+    def test_last_match_wins(self):
+        # Grand total should beat subtotal
+        text = 'Subtotal 18.00\nTax 1.50\nTotal 19.50'
+        assert _extract_amount_from_text(text) == 19.50
+
+    def test_empty_text(self):
+        assert _extract_amount_from_text('') == 0.0
+
+    def test_no_total_line(self):
+        assert _extract_amount_from_text('No price here') == 0.0
+
+    def test_comma_in_amount(self):
+        assert _extract_amount_from_text('Grand Total: $1,234.56') == 1234.56
+
+
+class TestExtractDate:
+    def test_iso_format(self):
+        assert _extract_date_from_text('Date: 2026-05-09') == '2026-05-09'
+
+    def test_slash_iso(self):
+        assert _extract_date_from_text('2026/05/09') == '2026-05-09'
+
+    def test_us_format(self):
+        assert _extract_date_from_text('05/09/2026') == '2026-05-09'
+
+    def test_us_short_year(self):
+        assert _extract_date_from_text('05/09/26') == '2026-05-09'
+
+    def test_no_date(self):
+        assert _extract_date_from_text('No date here') is None
+
+    def test_empty(self):
+        assert _extract_date_from_text('') is None
+
+
+# ---------------------------------------------------------------------------
+# _parse_receipt_text — combined extraction
 # ---------------------------------------------------------------------------
 
 @pytest.mark.asyncio
-async def test_parse_plain_ocr_text_uses_llm():
-    """Plain OCR text should go through the LLM extraction path."""
+async def test_parse_plain_ocr_text_uses_llm_for_vendor():
+    """Regex extracts amount; LLM called only for vendor + product_name."""
     agent = _make_agent()
     llm_resp = MagicMock()
-    llm_resp.content = '{"vendor":"Acme","amount":9.99,"date":"2026-05-09","time":null,"product_name":"Meals"}'
+    # LLM now only returns vendor + product_name
+    llm_resp.content = '{"vendor":"Acme","product_name":"Meals"}'
     agent._llm.submit = AsyncMock(return_value=llm_resp)
 
     result = await agent._parse_receipt_text(
@@ -439,10 +497,44 @@ async def test_parse_plain_ocr_text_uses_llm():
         expense_products=[{'id': 1, 'name': 'Meals'}],
     )
     assert result['vendor'] == 'Acme'
-    assert result['amount'] == 9.99
+    assert result['amount'] == 9.99   # from regex, not LLM
+    assert result['product_name'] == 'Meals'
     agent._llm.submit.assert_called_once()
 
 
+@pytest.mark.asyncio
+async def test_parse_date_hint_overrides_ocr_date():
+    """date_hint from filename must be used; LLM date should be ignored."""
+    agent = _make_agent()
+    llm_resp = MagicMock()
+    llm_resp.content = '{"vendor":"Shell","product_name":"Fuel"}'
+    agent._llm.submit = AsyncMock(return_value=llm_resp)
+
+    result = await agent._parse_receipt_text(
+        'Shell Gas\n05/09/2021\nTotal: $45.00', 'shell.jpg',
+        date_hint='2026-05-09',
+    )
+    assert result['date'] == '2026-05-09'   # filename timestamp wins
+    assert result['amount'] == 45.00
+
+
+@pytest.mark.asyncio
+async def test_parse_ocr_failed_skips_llm_amount():
+    """When OCR fails, amount=0 and date comes from hint or today."""
+    agent = _make_agent()
+    llm_resp = MagicMock()
+    llm_resp.content = '{"vendor":"","product_name":"Meals"}'
+    agent._llm.submit = AsyncMock(return_value=llm_resp)
+
+    result = await agent._parse_receipt_text(
+        '[Image: broken.jpg — OCR failed]', 'broken.jpg',
+        date_hint='2026-05-10',
+        expense_products=[{'id': 1, 'name': 'Meals'}],
+    )
+    assert result['amount'] == 0.0
+    assert result['date'] == '2026-05-10'
+
+
 # ---------------------------------------------------------------------------
 # parse_upload — receipt_parser.py
 # ---------------------------------------------------------------------------