Checkpoint cursors are known to be broken in durable history since they no longer represent a snapshot as of a particular time. Also, the cells now contain time pairs that influence visibility.
We should decide whether we can support reading directly from a checkpoint. If not, we should retire that functionality in the API and remove any associated testing. We've disabled some testing of checkpoint cursors in the wt4333_handle_locks test (just grep for FIXME) so we should either remove or re-enable this testing as part of this ticket.
haseeb.bokhari understands the problems involved with supporting checkpoint cursors so he is a good person to talk to before getting started with this work.