[HIGH] Aim 3.7: Intermediate negative-set sizes

From EXPERIMENT_QUEUE.md — Planned (run next)

Train with 0%, 25%, 50%, 75%, 100% of non-source personas in negative set.

• All A1 hyperparameters (lr=5e-5, r=16, 1 epoch), only vary negative set fraction
• Maps the emergence of the distance gradient as a function of contrastive pressure
• Compute: ~10 GPU-hours (5 conditions)
• Pod: any (parallelizable across multiple GPUs)