Plain text in the HTML header (think of <title>text</title>) is collected as usual. The decision to discard it is triggered by the first withspace following. In the usual case of the <body> no preceeded by any whitespace this means that by the point the first whitespace is encountered we're not in "discard mode" anymore and therefore incorrectly render the text.
Explicitly calling do_word() just before leaving "discard mode" discards all text collected so far.
957ac63a7c