I suspect one of two reasons for the differences: ...

Neither of those is the case.

I've verified (by sending the request through a proxy) that the content is sent UTF-8 encoded (i.e. no %-encoding).  And my Encode::decode also behaves normally (i.e. it would die with "Cannot decode string with wide characters").  This is, however, irrelevant, because CGI.pm has code to prevent double-decoding:

sub _decode_utf8 {
    my ($self, $val) = @_;

    if (Encode::is_utf8($val)) {
        return $val;
    }
    else {
        return Encode::decode(utf8 => $val);
    }
}
[download]

This sufficiently explains the behavior I observed and reported (for the input side).