Reproduced with the following page:
http://www.okaidi.fr/vetements-fille/?produits-par-page=1500
The _reader.reset() call in HtmlTokenizer.containsEndCData will produce a IOException with Mark invalid because we read too much from the input.
Reading one less char (read < MAX_BUFFER_SIZE - 1) will solve the issue.
Thanks Anthony, I'll look into it!
Good spot there - fixed in 2.23