Measured on run 19: three consecutive seeder/reseat sweeps returned bare "no legal pose anywhere on the board" for SW17+SW34+REF_PUCK_R (the third was a complete null — a passthrough board written with all three refs unmoved). When the question was finally asked in scoped form, the engine answered precisely: with D14 in place, 0 legal thumb-pocket poses; with D14 lifted, 46. With D31: 0 → 32. One eviction each and both switches seated, which closed the run.
Suggested fix: when a sweep ends at 0 legal poses for a ref, run a bounded second pass that lifts each nearby incumbent in turn and counts poses freed; return no_pose_blockers: {ref: {blocker: poses_freed}} in the JSON summary and print the top entries. A verdict that names its blockers is the next move; a bare verdict is a dead end that cost this run one wasted invocation and arguably the whole hand-seeding detour.
Full evidence: LEARNINGS.md finding 2 (link above).
Measured on run 19: three consecutive seeder/reseat sweeps returned bare "no legal pose anywhere on the board" for SW17+SW34+REF_PUCK_R (the third was a complete null — a passthrough board written with all three refs unmoved). When the question was finally asked in scoped form, the engine answered precisely: with D14 in place, 0 legal thumb-pocket poses; with D14 lifted, 46. With D31: 0 → 32. One eviction each and both switches seated, which closed the run.
Suggested fix: when a sweep ends at 0 legal poses for a ref, run a bounded second pass that lifts each nearby incumbent in turn and counts poses freed; return
no_pose_blockers: {ref: {blocker: poses_freed}}in the JSON summary and print the top entries. A verdict that names its blockers is the next move; a bare verdict is a dead end that cost this run one wasted invocation and arguably the whole hand-seeding detour.Full evidence: LEARNINGS.md finding 2 (link above).