Voilà la réponse, je cite :

byte order mark

Also known as BOM.

Name given to the Unicode character U+FEFF when used at the beginning of a Unicode byte stream. This invisible character generally know as ZERO WIDTH NO-BREAK SPACE (ZWNBSP) serves to identify unambiguously the Unicode transformation form used (and especially the byte order) for the stream. Indeed U+FFFE is a noncharacter so there is no risk of misinterpretation.

The following represents the byte signature of the character U+FEFF with the various Unicode Transformation Forms:
Bytes Encoding
00 00 FE FF UTF-32, big-endian
FF FE 00 00 UTF-32, little-endian
FE FF UTF-16, big-endian

FF FE UTF-16, little-endian
EF BB BF UTF-8

Historically the ZWNBSP was also used to indicate non-breaking but this use is now deprecated and replaced by the character U+2060 for that purpose.
"


http://p2p.wrox.com/archive/xslt/2002-11/79.asp
"
I suspect you edited the stylesheet using a text editor that inserts the
BOM at the start of the file. For example, I think Windows Notepad does
this.

Some XML parsers accept a BOM at the start of a UTF-8 file, others
don't. For example, the Crimson parser that comes with JDK 1.4 doesn't.

You basically need to find a text editor and an XML parser that agree
with each other about whether or not to use a byte order mark.
Alternatively, use a different encoding for the stylesheet, e.g.
iso-8859-1.
"

The code above found these strange bytes while expecting '<', and crashed of course.