Real-agent orchestration, replayed. At every junction an agent dissolves fully into one child per unexplored corridor; the first to reach the exit triggers a global halt and the lattice collapses to one route. Two memory models race the same maze.
Generated by stage2_harness.py (heuristic brain — navigation mechanical, the four meta-decisions stubbed). The same maze (seed 42, braided so it has loops) is solved under both memory models. Notice the spawn counts: shared prunes redundancy, blind re-walks it. Flip the harness to --brain claude with an API key to replace the stubs with real Claude judgment at the four decision seams — same visualization, real agency.