This patch was accepted by rubinius (along with a few other rubinius-specific fixes).
rubinius/rubinius@57068d0

So...was this a bug in MRI's date.rb? We largely ship the same stdlib code and don't like to diverge without good reason.

I do see that MRI's native date ext does the GREGORIAN version too:

/*
 * call-seq:
 *    t.to_date  ->  date
 *
 * Returns a Date object which denotes self.
 */
static VALUE
time_to_date(VALUE self)
{
    VALUE y, nth, ret;
    int ry, m, d;

    y = f_year(self);
    m = FIX2INT(f_mon(self));
    d = FIX2INT(f_mday(self));

    decode_year(y, -1, &nth, &ry);

    ret = d_simple_new_internal(cDate,
                nth, 0,
                GREGORIAN,
                ry, m, d,
                HAVE_CIVIL);
    {
    get_d1(ret);
    set_sg(dat, DEFAULT_SG);
    }
    return ret;
}

I went ahead and merged it. I think I understand that this was incorrect behavior inherited from 1.8, fixed in 1.9.2 and then ported into C for 1.9.3. Perhaps we should be incorporating 1.9.2's date.rb, if we're not already?

The 1.9 date library is unusual in so far as it is not loaded with the core, but it has a native implementation. Is anything like that done for JRuby? Are there any core classes which are extended by native standard libraries?