A write can execute too early compared to already scheduled reads

There is no problem with transaction ordering, and reads can not start too early. The problem is that the rule: "A transaction cannot start until all reads that executed before it in capture, have executed in replay" is not implemented.
From a scheduling perspective many reads can be scheduled to run in one thread (reading off of its queue), then if a transaction is scheduled to run on another thread, because all its dependant transactions have executed, it can be executed before the read thread has processed all its queued reads. Requires a lot of parallel sessions a fast machine, and full speed up replay (-s 0) to be reproducible.