Fix character references ending in raw nodes

Closes GH-15.

Author: wooorm

lib/index.js

@@ -28,17 +28,21 @@
  *   Way too simple, but works for us.
  * @property {HiddenPreprocessor} preprocessor
  * @property {(value: string) => void} write
+ * @property {() => number} _consume
  * @property {Array.<HiddenToken>} tokenQueue
  * @property {string} state
  * @property {string} returnState
  * @property {number} charRefCode
  * @property {Array.<number>} tempBuff
+ * @property {Function} _flushCodePointsConsumedAsCharacterReference
  * @property {string} lastStartTagName
  * @property {number} consumedAfterSnapshot
  * @property {boolean} active
  * @property {HiddenToken|undefined} currentCharacterToken
  * @property {HiddenToken|undefined} currentToken
  * @property {unknown} currentAttr
+ * @property {Function} NAMED_CHARACTER_REFERENCE_STATE
+ * @property {Function} NUMERIC_CHARACTER_REFERENCE_END_STATE
  *
  * @typedef {Object.<string, unknown> & {location: P5Location}} HiddenToken
  *
@@ -359,7 +363,20 @@ export const raw =
         // See the code for `parse` and `parseFragment`:
         // See: <https://github.com/inikulin/parse5/blob/9c683e1/packages/parse5/lib/parser/index.js#L371>.
         tokenizer.write(node.value)
-        parser._runParsingLoop(undefined)
+        parser._runParsingLoop(null)
+
+        // Character references hang, so if we ended there, we need to flush
+        // those too.
+        // We reset the preprocessor as if the document ends here.
+        // Then one single call to the relevant state does the trick, parse5
+        // consumes the whole token.
+        if (
+          tokenizer.state === 'NAMED_CHARACTER_REFERENCE_STATE' ||
+          tokenizer.state === 'NUMERIC_CHARACTER_REFERENCE_END_STATE'
+        ) {
+          preprocessor.lastChunkWritten = true
+          tokenizer[tokenizer.state](tokenizer._consume())
+        }
 
         // Process final characters if they’re still there after hibernating.
         // Similar to:

test.js

@@ -153,6 +153,24 @@ test('raw', (t) => {
     'should pass raw nodes through even after textarea (#3)'
   )
 
+  t.deepEqual(
+    raw(u('root', [u('raw', '{and}')])),
+    u('root', {data: {quirksMode: false}}, [u('text', '{and}')]),
+    'should pass character references through (decimal)'
+  )
+
+  t.deepEqual(
+    raw(u('root', [u('raw', '<and>')])),
+    u('root', {data: {quirksMode: false}}, [u('text', '<and>')]),
+    'should pass character references through (named)'
+  )
+
+  t.deepEqual(
+    raw(u('root', [u('raw', '&#x7b;and&#x7d;')])),
+    u('root', {data: {quirksMode: false}}, [u('text', '{and}')]),
+    'should pass character references through (hexadecimal)'
+  )
+
   t.deepEqual(
     raw(u('root', [u('raw', '<template>a<b></b>c</template>')])),
     u('root', {data: {quirksMode: false}}, [