Creating an issue for follow up on gh-29129. Also ping @mhvk because I am sure I'll forget things right now.

Some of these are very concrete, others a bit less so:

• There is now the _FLAG which part which is necessary for user dtypes. We should document at least that flag.
• I am not quite sure, but related, we may need to document the new flag on the context (likely is already)
• The design is in a way so that we could thread this cast to all places easily. Explicitly np.can_cast also.
  • @mattip wants to not really do this unless there is a clear user.
  • We might need this, but I am not quite sure. The problem is projects like cupy or so that might rely on e.g. can_cast.
  • Other places, where I think it just clearly makes sense is e.g. concatenate with dtype=.
• There is a clear gap that equiv casts are always same_value casts. But for existing dtypes this isn't the case. It would be good to explore if we can automatically set the flag for the dtype.
  • Maybe during registration or in MinCastSafety, or... ?