Set up your input stimuli (CK, D) as vpulse sources and make
the D source delay a variable.
Step the D delay such that, across iterations, it slides past
the CK active edge (here, the latch edge that makes it hold).
You will see there is a point where the output (Q) flips from
following the D input's change, to holding its pre-change
value.
Now if you subdivide between the last delay that followed,
and the first delay value that held, step delay at finer grain,
you should see more fine structure - hold, followed by a
period of increasing delay, then true metastability, then it
snaps to following.
For purposes of timing analysis you do not only care about
true metastability, but the "goalpost" span where latch
delay exceeds timing model limits. One end of this range is
your setup time and the other is your hold time.