fix(json): reject raw unpaired surrogates in string lexing instead of aborting

json/lex_string.mbt

@@ -17,7 +17,10 @@ fn ParseContext::lex_string(ctx : ParseContext) -> String raise ParseError {
   let string_start = ctx.offset
   // Fast path for ordinary strings: scan raw UTF-16 code units and materialize
   // the slice directly when there are no escapes or control characters.
-  for i in string_start..<ctx.end_offset {
+  for i = string_start {
+    if i >= ctx.end_offset {
+      break
+    }
     let c = ctx.input.unsafe_get(i)
     if c == '"' {
       ctx.offset = i + 1
@@ -29,7 +32,23 @@ fn ParseContext::lex_string(ctx : ParseContext) -> String raise ParseError {
       // \t) are invalid inside a JSON string.
       ctx.offset = i + 1
       ctx.invalid_char(shift=-1)
+    } else if c.is_leading_surrogate() {
+      if i + 1 < ctx.end_offset &&
+        ctx.input.unsafe_get(i + 1).is_trailing_surrogate() {
+        continue i + 2
+      }
+      // MoonBit strings stay Unicode well-formed, so a raw unpaired
+      // surrogate must be rejected rather than smuggled into the parsed
+      // string. (Well-formed input cannot contain one, but unsafe code can
+      // manufacture such a String; a clean error beats undefined behavior.)
+      ctx.offset = i + 1
+      ctx.invalid_char(shift=-1)
+    } else if c.is_trailing_surrogate() {
+      // A bare trailing surrogate can never start a surrogate pair.
+      ctx.offset = i + 1
+      ctx.invalid_char(shift=-1)
     }
+    continue i + 1
   }
   raise InvalidEof
 }
@@ -39,8 +58,14 @@ fn ParseContext::lex_string_slow(ctx : ParseContext) -> String raise ParseError
   let buf = StringBuilder()
   let mut start = ctx.offset
   fn flush(end : Int) {
-    if start > 0 && end > start {
-      buf.write_view(ctx.input[start:end])
+    if end > start {
+      // `view(start_offset~, end_offset~)` only bounds-checks. The checked
+      // `ctx.input[start:end]` would abort on a trailing surrogate at a
+      // slice boundary, which a raw lone surrogate inside the string could
+      // place there; unpaired surrogates now raise a ParseError before any
+      // flush spans them, and the unchecked slice keeps this loop's
+      // totality independent of that validation order.
+      buf.write_view(ctx.input.view(start_offset=start, end_offset=end))
     }
   }
 
@@ -70,12 +95,20 @@ fn ParseContext::lex_string_slow(ctx : ParseContext) -> String raise ParseError
         }
         start = ctx.offset
       }
-      Some(ch) =>
-        if ch.to_int() < 32 {
+      Some(ch) => {
+        let code = ch.to_int()
+        if code < 32 {
+          ctx.invalid_char(shift=-1)
+        } else if code is (0xD800..=0xDFFF) {
+          // `read_char` only yields a surrogate-range value when the raw
+          // code unit is unpaired; keep parsed strings Unicode well-formed
+          // by rejecting it (previously this aborted the process when a
+          // later flush sliced across the surrogate).
           ctx.invalid_char(shift=-1)
         } else {
           continue
         }
+      }
       None => raise InvalidEof
     }
   }

json/lex_string_test.mbt

@@ -89,3 +89,37 @@ test "lex_hex_digits accepts all hex digit ranges" {
     ),
   )
 }
+
+///|
+/// Regression for #4049. MoonBit strings must stay Unicode well-formed, so
+/// the string lexer rejects raw unpaired surrogates with a ParseError.
+/// Previously the escape-free fast path silently accepted them (producing
+/// ill-formed strings), and combining one with an escape sequence aborted
+/// the process: the slow path's `flush` sliced with the checked
+/// `[start:end]`, which panics when a slice boundary lands on a trailing
+/// surrogate. Well-formed input cannot contain raw lone surrogates, but
+/// unsafe code can manufacture such a String; a clean error beats an abort.
+test "raw unpaired surrogates are rejected with a clean parse error" {
+  let lone_high = String::from_array([(0xD800).unsafe_to_char()])
+  let lone_low = String::from_array([(0xDC00).unsafe_to_char()])
+  // Escape-free strings (fast path; used to be silently accepted).
+  assert_false(@json.valid("\"" + lone_high + "\""))
+  assert_false(@json.valid("\"" + lone_low + "\""))
+  // A reversed pair (low then high) is two unpaired surrogates.
+  assert_false(@json.valid("\"" + lone_low + lone_high + "\""))
+  // Next to escapes (slow path; used to abort the process).
+  assert_false(@json.valid("\"" + lone_low + "\\n\""))
+  assert_false(@json.valid("\"\\n" + lone_low + "\""))
+  assert_false(@json.valid("\"" + lone_high + "\\t\""))
+  // The failure is the documented ParseError, not an abort.
+  debug_inspect(
+    expect_parse_error("\"" + lone_low + "\"", "expected InvalidChar"),
+    content=(
+      #|InvalidChar({ line: 1, column: 1 }, '�')
+    ),
+  )
+  // Well-formed surrogate pairs still parse, raw or next to escapes.
+  assert_true(@json.parse("\"\u{1F600}\"") == Json::string("\u{1F600}"))
+  let json = Json::string("\u{10FFFF}\\\u{1F600}\n")
+  assert_true(@json.parse(json.stringify()) == json)
+}