feat(tn): en-dash kept literal + detokenized (punctuation 42→43)

src/lib.rs

@@ -1062,7 +1062,10 @@ fn is_split_punct(c: char) -> bool {
 /// never occur inside glued numeric/semiotic forms (unlike `.`/`:`/`,`), so
 /// splitting them cannot break decimals, times, or IPs.
 fn is_interior_hard(c: char) -> bool {
-    matches!(c, '!' | '?' | '(' | ')' | '[' | ']' | '{' | '}' | '\'')
+    matches!(
+        c,
+        '!' | '?' | '(' | ')' | '[' | ']' | '{' | '}' | '\'' | '\u{2013}'
+    )
 }
 
 /// Fold NeMo's double-backtick quotes to a straight double quote so the
@@ -1242,7 +1245,10 @@ where
 /// "hello" → "one! hello"), matching NeMo's punctuation post-processing.
 fn push_detokenized(out: &mut String, sep: &str, text: &str) {
     let glue_after_close = sep.is_empty()
-        && matches!(out.chars().last(), Some('!' | '?' | ')' | ']' | '}'))
+        && matches!(
+            out.chars().last(),
+            Some('!' | '?' | ')' | ']' | '}' | '\u{2013}')
+        )
         && text
             .chars()
             .next()

src/tn/en/range.rs

@@ -32,12 +32,11 @@ pub fn parse(input: &str) -> Option<String> {
         }
     }
 
-    // Hyphen or en-dash range.
-    for sep in ['-', '\u{2013}'] {
-        if let Some((a, b)) = split_binary(t, sep) {
-            if let Some(result) = range_words(a, b) {
-                return Some(result);
-            }
+    // Hyphen range ("1980-1986" → "… to …"). An en-dash is kept literal by the
+    // tokenizer/detokenizer, so it is not treated as a range here.
+    if let Some((a, b)) = split_binary(t, '-') {
+        if let Some(result) = range_words(a, b) {
+            return Some(result);
         }
     }
 

tests/parity_baseline.tsv

@@ -58,7 +58,7 @@ en	tn	measure	14	21
 en	tn	money	71	71
 en	tn	normalize_with_audio	40	58
 en	tn	ordinal	27	27
-en	tn	punctuation	42	63
+en	tn	punctuation	43	63
 en	tn	punctuation_match_input	6	13
 en	tn	range	20	20
 en	tn	roman	4	4