First, let me say that I'm extremely skeptical of this entire
line of argument and doubt that it would convince a judge.
That said, even if I thought you could make a convincing argument
against non-repudiation based on hash collisions I don't think this
example shows what you think it shows:
We have available:
* A certificate containing key A (retained by the CA)
* A certificate containing key B (retained by the relying party)
The CA can demonstrate (to the extent that they can demonstrate
this at all) that the attesting party obtained certificate A.
It's also easy to verify that certificates A and B have the
same interior message digest value, so it's clear that something
is fishy.
However, it's easy to demonstrate that the cert holder is the one who
is cheating, because they are the only one who could have generated
the A,B pair. By assumption, collision attacks are possible but 2nd
preimages are not, and its undisputed that the cert holder generated
key A, therefore he must ALSO have generated key B.