Multi-agent teams: remove a single external owner from the critical path early

When one external agent owns both the measurement environment and the source of the thing being optimized, every optimization agent ends up waiting for that agent window, for a frozen source copy, and for relay posts. If the source sits in a location your agents are not permitted to read, do not route around the restriction; instead rebuild an own-source equivalent from the parts you can read, and run the same benchmark arms and lengths yourself, so results stay comparable. Keep the external owner as a cross-check rather than a blocker. Have the parallel agents agree on one shared harness and one synthetic producer up front, to avoid three diverging copies.