The schema_disagg_abort generator flushes every pending publish before it emits the step-down marker, so the step-down boundary's publish rules never run against a live workload. The flush is a workaround for ending a term with unpublished operations (FIXME-WT-18272, FIXME-WT-18284) and it also hides the checkpoint panic a deferred drop can hit (WT-18324).

Rework the harness so the boundary rules are exercised with the boundary set:

• Publish the term's remaining unpublished operations behind the step-down marker, at epochs reserved at or below the boundary. The step-down checkpoint runs synchronously at the marker, so the only room for these publishes is between declaring the boundary and advancing the stable frontier to it: the generator bumps the timestamp counter past the reserved range, stamps the marker with the boundary, and carries the publish count on the marker event so the reader can drain exactly that many publishes before it advances the frontier and takes the checkpoint.
• The reservation must be exact-count: an unused reserved timestamp holes the completed-timestamp window and stalls the frontier permanently.
• Remove the pre-marker flush. A term can then end holding unpublished drops, which surfaces the WT-18324 checkpoint panic in the window-drop shape instead of working around it.

A working implementation exists and was parked out of the WT-18476 branch pending review by someone familiar with the harness. Note the harness has since gained the epoch-less mode and a new set_ts signature, so the parked diff needs rebasing.