fix(failure): read a trailing context limit after a multi-byte character

crates/tinyinference-llm/src/failure/context_limit.rs

@@ -132,9 +132,14 @@ fn first_number(text: &str, window: usize) -> Option<u64> {
 /// window.
 fn trailing_number(text: &str) -> Option<u64> {
     let trimmed = text.trim_end();
+    // Step past the whole character that precedes the digits: it can be
+    // multi-byte (a no-break space, a CJK character), and `index + 1` would
+    // land inside it.
     let begin = trimmed
-        .rfind(|character: char| !(character.is_ascii_digit() || character == ','))
-        .map_or(0, |index| index + 1);
+        .char_indices()
+        .rev()
+        .find(|&(_, character)| !(character.is_ascii_digit() || character == ','))
+        .map_or(0, |(index, character)| index + character.len_utf8());
     let digits = trimmed[begin..].trim_start_matches(',');
     if digits.is_empty() {
         return None;

crates/tinyinference-llm/src/failure/context_limit_tests.rs

@@ -43,6 +43,16 @@ fn parses_mistral_trailing_limit() {
     assert_eq!(parse_context_limit_from_error(message), Some(32_768));
 }
 
+#[test]
+fn parses_trailing_limit_after_a_multibyte_character() {
+    // A no-break space or a CJK character right before the number used to
+    // panic: the start of the digits was taken as one byte past it.
+    let message = "too large for model with\u{a0}32768 maximum context length";
+    assert_eq!(parse_context_limit_from_error(message), Some(32_768));
+    let message = "输入过长，模型32768 maximum context length";
+    assert_eq!(parse_context_limit_from_error(message), Some(32_768));
+}
+
 #[test]
 fn parses_gemini_allowed_maximum() {
     let message = "The input token count (1200000) exceeds the maximum number of tokens \