I have more experience with fold-and-dock approaches (e.g., AlphaFold, ESMFold, Boltz), and with protein-protein interaction (PPI) data, but in all this is a pretty common outcome.

I'm not familiar with tools used to assess protein-ligand docking, but I am sure there are a few you can add that score the interaction in some way or another (e.g., pDockQ for PPI). Alternatively, there are some libraries that exist specifically for this, such as MDanalysis and pymol libraries in python.

I usually start with a simple filter to remove nonsense models such as those with clashes. In your example, you could quickly filter any model where the ligand and the zinc are far apart.

For my results, graphs, and text, can I:
Use the binding affinity value from pose 3 (lowest energy),
But use pose 10 only for visualization

Do not do this. You are misrepresenting the data, and this can easily be misinterpreted unless you are clear and explicit. You can write it up in a way that shows the first n poses appear to be nonsense and give your reason, and as such pose 10 appears to be the most realistic.

For data analysis (results, interpretation, graphs etc), can I use the lowest docking energy for each protein–ligand pair, because it is not practical to visually inspect every pose for all compounds.

You'll likely need some way other than just binding free energy to rank these. You can also write a simple pymol or similar script that visualises each pose in a png, for example.