fix: improve UTF-16 detection logic in is_binary function

Co-authored-by: aider (openai/andrew/openrouter/qwen/qwen3-coder) <aider@aider.chat>

src/common/is_binary.rs

@@ -150,18 +150,30 @@ fn looks_like_utf16(data: &[u8]) -> bool {
         return false;
     }
     
-    let mut zero_count = 0;
+    // Check if it could be UTF-16 by looking at null patterns
-    let pairs = data.len() / 2;
+    let mut null_pairs = 0;
+    let max_checks = std::cmp::min(data.len() / 2, 50); // Check up to 50 character pairs
     
-    // Check if every other byte is zero (indicating UTF-16)
+    for i in 0..max_checks {
-    for i in 0..pairs {
         if data[i * 2 + 1] == 0 {
-            zero_count += 1;
+            null_pairs += 1;
         }
     }
     
-    // If more than 50% of odd positions are zero, might be UTF-16
+    // If most high bytes are zero, it's likely UTF-16
-    zero_count as f64 / pairs as f64 > 0.5
+    if max_checks > 0 && null_pairs as f64 / max_checks as f64 > 0.7 {
+        return true;
+    }
+    
+    // Also check the reverse pattern (little-endian UTF-16)
+    let mut null_pairs_reverse = 0;
+    for i in 0..max_checks {
+        if i * 2 + 1 < data.len() && data[i * 2] == 0 {
+            null_pairs_reverse += 1;
+        }
+    }
+    
+    null_pairs_reverse as f64 / max_checks as f64 > 0.7
 }
 
 /// Check if data looks like a TAR archive