ensemble_mean_q_buf and ensemble_var_q_buf were allocated as batch_size * total_actions (12), but the kernel writes batch_size * num_atoms (51) elements. 2977 OOB write errors. Fixed: allocate batch_size * num_atoms. compute-sanitizer: 0 errors. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>