fix: broaden JSON answer extraction for model formatting variations

fix: wrap toxigen samples_start past dataset end instead of crashing

feat: ethical refusal as scored dimension, not failure

feat: generative toxigen eval with silent exit detection

fix: Pillow in PEP 723 deps + capture lighteval stderr on failure

feat: add quant field — (name, type, quant) is the composite target key

fix: capability probe promotes both backends independently

feat: gguf wrapper + (name, type) composite target key

refactor: partition by type (mlx|gguf), drop owner hostname field

feat: initial LEM-Eval scaffold — target-driven 8-PAC runner