PrismaQuant: GPTQ + closed-form scale_sweep polish (replaces act_round). Measured ~67% per-Linear output-MSE reduction vs prior pipeline in bake-off.

PrismaQuant final: visual NVFP4 (108/110 DP-placed), lm_head BF16 (vLLM ParallelLMHead runtime limit), pos_embed excluded

PrismaQuant Phase 2 incremental: body AWQ-safe GPTQ+act-round, visual Fisher-allocated (108/111 Linears via per-Linear DP)