What this is:        redux2 EDGE variant -- max abliteration preservation under 4 GB
Trial:               Best Pareto trial from redux2 (50-trial Heretic + projection_pass_v2)
Quant scheme:
  tier1 (8-bit affine, group 64) : mlp.down_proj on ALL 36 layers
                                    (residual-stream writer, heavily modified
                                     by projection_pass_v2)
  tier2 (6-bit affine, group 64) : self_attn.o_proj on LATE layers (12-35) only
                                    (the layers projection_pass_v2 actually
                                     touched -- their weights drifted enough
                                     from ternary to tolerate affine quant)
  rest                            : 2-bit ternary (zero-preserving custom scheme)
Why "Lo6":           "Late-layer o_proj at 6-bit". Restricting to layers 12-35
                     avoids the redux1-mixed-do6 failure mode where ALL 36
                     attn.o_proj at 6-bit broke the model (early layers are
                     too-close-to-pure-ternary; 6-bit affine has no zero level).
Config:              rope_theta=1000000.0, rope_scaling=yarn (MLX-LM compatible)