Refactor documentation and remove commented-out code in training adapter

antipasto/peft_utils/adapter_scaling.py

@@ -4,7 +4,7 @@ Adapter steering for contrastive training with proper gradient flow.
 For AntiPaSTO: Sets `antipasto_alpha` directly on each layer. The Cayley rotation
 transform satisfies R(-α) = R(α)^(-1), so a single adapter handles both steering directions.
 
-Key insight: PyTorch's autograd tracks tensor references in the computation graph,
+PyTorch's autograd tracks tensor references in the computation graph,
 not module attributes. So we can:
 1. Replace `module.weight` with `weight * coeff` (graph stores ref to original param)
 2. Run forward pass

antipasto/train/train_adapter.py

@@ -1968,7 +1968,6 @@ def train_model(config: TrainingConfig):
                 f"Results for method: {method} [logratio * label -> nat's toward label]\n{df_res_pv[method].head(5).round(4)}\n"
             )
 
-    # Generate comprehensive metrics (both text and markdown)
     md_table, tables_dict, main_score = format_main_results_table(
         df_res_wlabels,  config=config
     )