Fix grammar JSON spacing bug

src/nlp/grammar/grammar_rules.py

@@ -718,7 +718,14 @@ class ArabicGrammarGuard:
         """Apply all grammar rules to model output."""
         text = self.preserve_numbers(original_text, generated_text)
         
-        # ── Fix Hallucinated Subject Gender & Protect Structured Data ──
+        # ── Fix T5 Spacing Hallucinations & Protect Structured Data ──
+        # T5 often adds spaces around structural brackets. Fix them before splitting.
+        for bracket in ['{', '}', '[', ']', '<', '>']:
+            if f' {bracket}' not in original_text:
+                text = text.replace(f' {bracket}', bracket)
+            if f'{bracket} ' not in original_text:
+                text = text.replace(f'{bracket} ', bracket)
+                
         orig_words = original_text.split()
         corr_words = text.split()
         if len(orig_words) == len(corr_words):
@@ -728,8 +735,8 @@ class ArabicGrammarGuard:
                 o_clean = o.rstrip('.,،؛;:!؟?()[]{}«»"\'…')
                 c_clean = c.rstrip('.,،؛;:!؟?()[]{}«»"\'…')
                 
-                # Protect structured data (English, JSON, Hashtags, Code)
-                if re.search(r'[a-zA-Z]|\{.*\}|\[.*\]|<.*>|#\S+|@\S+', o):
+                # Protect structured data (English, JSON, Code, URLs)
+                if re.search(r'[a-zA-Z]|\{|\[|<|#|@|://', o):
                     corr_words[i] = o
                     # Revert grammar hallucinations on adjacent Arabic words caused by structured data
                     if i > 0:

src/nlp/punctuation/punctuation_rules.py

@@ -167,7 +167,7 @@ def validate_punctuation_diff(diff: dict, full_text: str = '') -> bool:
 
     # ── Protect Structured Data (English, URLs, Emails, Hashtags, Code/JSON) ──
     # Block punctuation modifications near structured data unless it's a valid terminal punctuation
-    if re.search(r'[a-zA-Z]|\{.*\}|\[.*\]|<.*>|#\S+|@\S+', original):
+    if re.search(r'[a-zA-Z]|\{|\[|<|#|@|://', original):
         is_at_end = False
         if full_text and 'end' in diff:
             is_at_end = diff['end'] >= len(full_text) - 2
@@ -183,7 +183,7 @@ def validate_punctuation_diff(diff: dict, full_text: str = '') -> bool:
             return False
             
         # Block spacing corruptions in JSON/Code (e.g. {"name"} -> { "name" })
-        if re.search(r'\{.*\}|\[.*\]|<.*>|https?://', original):
+        if re.search(r'\{|\[|<|://', original):
             # Only allow if the ONLY change is appending a terminal mark at the very end
             if original != correction and not (is_at_end and correction.endswith(('.', '؟')) and correction[:-1].rstrip() == original.rstrip()):
                 logger.info(f"[PUNC-SAFETY] Blocked corruption of JSON/Code/URL: '{original}' -> '{correction}'")