Also, should "current_user_ns() != token->userns" perhaps be an error
condition, rather than a "fall back to init_ns" condition?

Yes, I've pointed this out before:

"Please enforce that in order to use a token the caller must be in the
same user namespace as the token as well. IOW, we don't want to yet make
it possible to use a token created in an ancestor user namespace to load
or attach bpf programs in a descendant user namespace. Let's be as
restrictive as we can: tokens are only valid within the user namespace
they were created in."

[1] Re: [PATCH v11 bpf-next 03/17] bpf: introduce BPF token object
    https://lore.kernel.org/r/20231130-katzen-anhand-7ad530f187da@brauner (local)

Again, none of this is a big deal. I do think you're dropping the LSM
error code on the floor, and are duplicating the "ns_capable()" vs
"capable()" logic as-is, but none of this is a deal breaker, just more
of my commentary on the patch and about the logic here.

And yeah, I don't exactly love how you say "ok, if there's a token and
it doesn't match, I'll not use it" rather than "if the token namespace
doesn't match, it's an error", but maybe there's some usability issue
here?