Probably in order to implement case-sensitive key matching when the Windows
registry itself is case-insensitive.

You may find the comment from the relevant method interesting, but I presume
that you should not rely on the details remaining constant:

===========
    * Converts value's or node's name to its Windows representation
    * as a byte-encoded string.
    * Two encodings, simple and altBase64 are used.
    * <p>
    * <i>Simple</i> encoding is used, if java string does not contain
    * any characters less, than 0x0020, or greater, than 0x007f.
    * Simple encoding adds "/" character to capital letters, i.e.
    * "A" is encoded as "/A". Character '\' is encoded as '//',
    * '/' is encoded as '\'.
    * The constructed string is converted to byte array by truncating the
    * highest byte and adding the terminating <tt>null</tt> character.
    * <p>
    * <i>altBase64</i>  encoding is used, if java string does contain at least
    * one character less, than 0x0020, or greater, than 0x007f.
    * This encoding is marked by setting first two bytes of the
    * Windows string to '/!'. The java name is then encoded using
    * byteArrayToAltBase64() method from
    * Base64 class.

===========

The Base64.byteArrayToAltBase64() method it refers to is from another
Sun-private class.  It appears to do roughly the same thing as "normal" Base64
encoding, but using a different set of encoded characters in order to avoid
uppercase letters.

   -- chris