Summary

During the implementation of WT-10570 (enable REMOVE in test/format predictable replay), a per-operation logging mechanism was added to help diagnose data mismatches between runs. The logging was removed before merging because it ended up adding significant complexity to already-complex test/format logic. This ticket tracks a future clean implementation of that logging if it proves useful.

Scope

If diagnostics are needed for future predictable replay debugging, implement a simpler logging mechanism that:

• Writes to a per-run file only (no stderr by default) — format is run interactively by the team
• Minimises state carried across the transaction loop (ideally no persistent flags between op selection and commit)
• Handles the retry path cleanly so re-seeded data_rnd on rollback doesn't produce duplicate or out-of-order log entries
• Is clearly guarded by GV(RUNS_PREDICTABLE_REPLAY) and has no impact on the normal non-replay code path

Acceptance criteria

• Given two diverging predictable replay runs, a diff of their log files immediately identifies the first divergent operation (lane, ts, keyno, table, ret)
• No output to stderr unless explicitly requested
• No increase in complexity of the ops loop rollback/retry paths