fix: vision OCR JSON failures — add format='json' and repair fallback

Three receipts per batch were failing with JSONDecodeError (e.g.
"Expecting ':' delimiter: line 1 column 90") because activeblue-chat
(llama3.2-vision) occasionally outputs near-JSON with trailing commas,
single-quoted strings, or unquoted keys.

Two-layer fix:
1. Add format='json' to the Ollama chat call — Ollama JSON mode forces
   syntactically valid output at the sampler level, eliminating most
   structural errors.
2. Add _repair_json() fallback that runs on any remaining JSONDecodeError:
   strips trailing commas, converts single→double quotes, and quotes
   unquoted keys. If repair succeeds, the result is re-serialised as
   canonical JSON before being returned.

Also re-serialise with json.dumps() on success so the fast path in
_parse_receipt_text always receives clean, canonical JSON regardless of
whitespace or key ordering in the model's original output.

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

agent_service/tools/receipt_parser.py

@@ -105,11 +105,30 @@ def _ocr_image_vision(data: bytes, filename: str, ollama_url: str, model: str) -
     Returns empty string on any failure so the caller falls back to Tesseract.
     """
     import json as _json
+    import re as _re
+
+    def _repair_json(s: str) -> str:
+        """Fix the most common LLM JSON formatting mistakes.
+
+        Handles:
+        - trailing commas before } or ]  →  {"a":1,}  becomes  {"a":1}
+        - single-quoted strings          →  {'a':'b'}  becomes  {"a":"b"}
+        - unquoted string keys           →  {a: "b"}   becomes  {"a": "b"}
+        """
+        # trailing commas
+        s = _re.sub(r',\s*([}\]])', r'\1', s)
+        # single-quoted strings (careful around apostrophes in values)
+        s = _re.sub(r"'([^']*)'", r'"\1"', s)
+        # unquoted keys: word characters before a colon
+        s = _re.sub(r'(?<!["\w])(\w+)\s*:', r'"\1":', s)
+        return s
+
     try:
         import ollama as _ollama
         client = _ollama.Client(host=ollama_url)
         response = client.chat(
             model=model,
+            format='json',   # Ollama JSON mode — forces syntactically valid output
             messages=[{
                 'role': 'user',
                 'content': (
@@ -148,12 +167,26 @@ def _ocr_image_vision(data: bytes, filename: str, ollama_url: str, model: str) -
                            filename)
             return ''
         json_str = raw[first:last + 1]
-        parsed = _json.loads(json_str)
+
+        # Parse — on failure attempt common repairs then retry once
+        try:
+            parsed = _json.loads(json_str)
+        except _json.JSONDecodeError as json_err:
+            repaired = _repair_json(json_str)
+            try:
+                parsed = _json.loads(repaired)
+                logger.debug('Vision OCR %s: JSON repaired successfully', filename)
+            except _json.JSONDecodeError:
+                logger.warning('Vision OCR %s: JSON parse failed (%s), falling back',
+                               filename, json_err)
+                return ''
+
         if 'amount' not in parsed:
             logger.warning('Vision OCR %s: JSON missing amount field, falling back', filename)
             return ''
         logger.debug('Vision OCR %s (%s): extracted JSON ok', filename, model)
-        return json_str
+        # Re-serialise so downstream always gets clean, canonical JSON
+        return _json.dumps(parsed)
     except ImportError:
         logger.warning('ollama package not installed — vision OCR unavailable for %s', filename)
         return ''