fix: block grammar كان→كأن + punctuation mid-word split (الدفت.ر)

src/app.py

@@ -1612,6 +1612,22 @@ def analyze_text():
                     stage_label = 'grammar'
                     if _is_spelling_only_change(orig_text, corr_text):
                         stage_label = 'spelling'
+
+                    # ── Directional blocks for grammar (mirrors spelling filter) ──
+                    # Prevents grammar from making meaning-changing corrections
+                    # like كان→كأن ("was" → "as if").
+                    _GRAMMAR_BLOCKS = {
+                        'كان': {'كأن'}, 'كأن': {'كان'},
+                        'هذه': {'هذة'}, 'هذا': {'هذة', 'هذه'},
+                        'إلى': {'على', 'علي'}, 'على': {'إلى', 'علي'},
+                        'لكن': {'لاكن'}, 'ذلك': {'ذالك'},
+                    }
+                    if corr_text in _GRAMMAR_BLOCKS.get(orig_text, set()):
+                        logger.info(
+                            f"[GRAMMAR] Blocked directional: '{orig_text}'→'{corr_text}'"
+                        )
+                        continue
+
                     ctx.add_patch(
                         stage_label, d['start'], d['end'],
                         corr_text, confidence=1.0
@@ -1674,6 +1690,20 @@ def analyze_text():
                             f"'{d.get('original','')}' \u2192 '{d.get('correction','')}' "
                             f"(locked by {owner}[{ls}:{le}])"
                         )
+                    # ── Mid-word split guard ──
+                    # Reject punctuation diffs where the original is NOT a complete
+                    # word — i.e., the character after the diff end is still Arabic.
+                    # This catches cases like الدفتر being split into الدفت.ر
+                    d_end = d['end']
+                    if d_end < len(ctx.current_text):
+                        next_ch = ctx.current_text[d_end]
+                        if '\u0600' <= next_ch <= '\u06FF':
+                            logger.info(
+                                f"[PUNC-SAFETY] Rejected mid-word split [{d['start']}:{d_end}] "
+                                f"'{d.get('original','')}' → '{d.get('correction','')}' "
+                                f"(next char '{next_ch}' is Arabic — word was split)"
+                            )
+                            continue
                     # Punctuation safety layer: reject non-punctuation changes
                     if not validate_punctuation_diff(d):
                         logger.info(

src/nlp/punctuation/punctuation_rules.py

@@ -137,5 +137,19 @@ def validate_punctuation_diff(diff: dict) -> bool:
     if punct_delta > MAX_PUNCT_DELTA:
         return False
 
+    # ── Rule 6: Reject mid-word punctuation insertion ──
+    # If the correction ends with a punctuation mark followed by nothing,
+    # but the original word is a PREFIX of a longer word in context,
+    # this indicates mid-word split (e.g. الدفت→الدفت. when word was الدفتر).
+    # Detect by checking if correction has punctuation NOT at word boundary.
+    for pc in ARABIC_PUNCT_CHARS:
+        if pc in correction:
+            # Check if punctuation is followed by an Arabic letter (mid-word)
+            idx = correction.find(pc)
+            if idx >= 0 and idx < len(correction) - 1:
+                next_char = correction[idx + 1]
+                if '\u0600' <= next_char <= '\u06FF':
+                    return False
+
     return True