fix(ocr+safety): stop document-content hallucinations from empty OCR

Three independent bugs combined to make Shira fabricate the content of
case-46-Friedman's appeal in production (CTS case rewritten as a knee
injury, wrong dates, wrong court file number, wrong %-of-disability —
written into save_memory/Task/Meeting on prod).

* api/services/ocr.py — ocr_docx_images now reads word/document.xml first
  (text content), then OCRs embedded images. Previously it only OCR'd the
  word/media/* images, so a text-heavy DOCX with one signature image
  returned only "[signature]" to the LLM. Verified against the same
  Friedman appeal: 16633 chars / 80 paragraphs / 0 XML noise.

* mcp_server/tools/document_tools.py — _ocr_fallback now refuses to wrap
  an empty/signature-only OCR result as success. If <80 useful chars or
  just "[signature]", returns an explicit failure that instructs the LLM
  not to describe / summarize the document.

* api/services/prompt_builder.py — TOOL_RULES adds the absolute rule
  "DOCUMENT FAITHFULNESS": never quote a document not literally in the
  tool result for THIS turn; never infer content from a file name; treat
  dates and case numbers as especially dangerous to invent.

Refs Task Master #5, #6, #7

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

mcp_server/tools/document_tools.py

@@ -97,6 +97,12 @@ def register_document_tools(
         except Exception as e:
             return fail(f"שגיאה בשמירת נתיב: {e}")
 
+    # Minimum length (in characters) of an extracted body that we consider
+    # "real text". Below this, the OCR almost certainly returned only chrome
+    # like "[signature]" or "תמונה ריקה" and the LLM has nothing to work with —
+    # so we MUST report failure explicitly, never wrap it as success.
+    MIN_USEFUL_TEXT_CHARS = 80
+
     async def _ocr_fallback(filePath: str, reason: str) -> str:
         """Fetch file bytes and OCR via Claude Vision."""
         from api.services.ocr import extract_text_from_bytes
@@ -115,10 +121,28 @@ def register_document_tools(
             text = await extract_text_from_bytes(
                 bytes_result["base64"], mime_type, file_name
             )
-            return ok(f'=== {file_name} (OCR via Vision) ===\n\n{text}')
         except Exception as e:
             return fail(f'שגיאה ב-OCR של "{file_name}": {e}')
 
+        # Guard against the hallucination trap: if OCR returned only a signature
+        # or a near-empty result, the LLM must NOT be allowed to invent content
+        # to fill the gap. Return an explicit failure so it tells the user it
+        # could not read the document.
+        body = (text or "").strip()
+        body_lower = body.lower()
+        looks_empty = (
+            len(body) < MIN_USEFUL_TEXT_CHARS
+            or body_lower in {"[signature]", "signature", "חתימה", "[חתימה]"}
+            or body.startswith("❌")
+        )
+        if looks_empty:
+            return fail(
+                f'לא הצלחתי לחלץ טקסט שמיש מ-"{file_name}" (התקבלו רק {len(body)} תווים, "{body[:60]}"). '
+                'אל תנסי לתאר או לסכם את המסמך — הטקסט לא הגיע אלייך. '
+                'הציעי למשתמש לפתוח את הקובץ ידנית או לנסות מסמך אחר.'
+            )
+        return ok(f'=== {file_name} (OCR via Vision) ===\n\n{body}')
+
     async def read_document(filePath: str) -> str:
         try:
             result = await crm.post("SmartAssistant/action/readDocument", {"filePath": filePath})