VRM V1.1 - Support Mapping Counterfactual Ablation

Evaluation-only diagnostic. No optimizer, gradient update, or checkpoint modification was used.

command: python -m vrm.train --mode v1_1_support_ablation --checkpoint /mnt/d/VRM/checkpoints/vrm_v1_1_best_train_hamming.pt --seed 20260627 --pytest-result pytest -q: 31 passed in 5.05s
checkpoint: /mnt/d/VRM/checkpoints/vrm_v1_1_best_train_hamming.pt
checkpoint_sha256_before: b86ba9fd645870e185ef1c54fbec98e5437aeac094d0e290e19bc388a5722370
checkpoint_sha256_after: b86ba9fd645870e185ef1c54fbec98e5437aeac094d0e290e19bc388a5722370
task_manifest: /mnt/d/VRM/artifacts/vrm_v1_1_capacity50/vrm_v1_1_task_manifest.json
train_episodes: 156
unseen_episodes: 160
train_output_deranged_eligible: 140
unseen_output_deranged_eligible: 148

Counterfactual metrics

split	condition	shape_match	target_color_acc	transformation_acc	exact	mean_hamming
train_50	NORMAL	0.935897	0.702464	0.247191	0.044872	47.134615
unseen_50	NORMAL	0.187500	0.678179	0.279529	0.000000	44.318750
train_50	OUTPUT_DERANGED_WITHIN_TASK	0.907143	0.698531	0.240575	0.050000	48.228571
unseen_50	OUTPUT_DERANGED_WITHIN_TASK	0.168919	0.673266	0.239695	0.000000	43.378378
train_50	PAIR_ORDER_SHUFFLED	0.814103	0.698256	0.241277	0.044872	47.801282
unseen_50	PAIR_ORDER_SHUFFLED	0.181250	0.679041	0.279693	0.000000	44.200000
train_50	OUTPUT_COLOR_PERMUTED	0.730769	0.695666	0.219791	0.051282	48.211538
unseen_50	OUTPUT_COLOR_PERMUTED	0.181250	0.659027	0.254370	0.000000	46.956250
train_50	SUPPORT_EVIDENCE_PATH_ZEROED	0.435897	0.682111	0.178987	0.000000	50.358974
unseen_50	SUPPORT_EVIDENCE_PATH_ZEROED	0.156250	0.662340	0.256004	0.000000	46.500000

Support-path zeroing is an out-of-distribution stress test and is supplementary to output derangement.