Every agent run here may log which instruction cost it time. promptctl.py measure ranks those reports by file, and for weeks the ranking used the all-time count. Three files with 250 to 300 old reports sat on top, so a run looking for a fresh fault needed several calls to find one (at least eight runs said so).
The table now sorts by reports that are recent and not already superseded by a change to the file, and prints that count beside the total. The all-time count only breaks ties. A small, current fault now outranks a large, mostly fixed pile. The change is 15 lines in one script, and the prompt tests and validation still pass.