Address issues 556 and 568

[contextablemark]

Note: This fix does not fully address #568; doing this will require some more extensive refactoring.

Addresses #568 and #556

Summary

• iterate unseen transcript messages in arrival order so tool results and user messages are handled without reordering
• allow _start_new_execution and the background runner to accept explicit message batches, ensuring only the intended messages are marked processed
• extend the tool-result flow tests to assert ordering and capture the batches passed through the execution helpers
• insure that assistant messages and tool requests don't get played back