Update docstrings to reflect LLM-only training pipeline

MockPromptOptimizer and mock mode now clearly document that they
use real LLM inference (Llama 3.1 8B) on both agent and customer
sides — only the prompt selection is "mock" (hand-picked vs GRPO).

https://claude.ai/code/session_01DPirJ78YYN4fJUvUFJ5D6V

layer1/grpo_trainer.py

@@ -275,10 +275,11 @@ class GRPOPromptTrainer:
 
 class MockPromptOptimizer:
     """
-    CPU-friendly optimizer for testing the pipeline end-to-end.
+    Evaluates hand-written candidate prompts using real LLM agent + customer.
 
-    Evaluates hand-written candidate prompts (simulating the evolution
-    from naive to optimized) and returns the best one.
+    Tests the pipeline end-to-end with actual Llama 3.1 8B on both sides.
+    The prompt selection is "mock" (hand-picked instead of GRPO-generated),
+    but evaluation uses real LLM inference to measure actual agent behavior.
     """
 
     CANDIDATE_PROMPTS = [

layer1/train.py

@@ -5,7 +5,7 @@ Usage:
     # Full GPU training (requires Colab/GPU + train deps)
     python -m layer1.train --mode train --steps 10
 
-    # CPU mock optimization (evaluates hand-written prompts)
+    # Mock optimization (evaluates hand-written prompts via real LLM agent)
     python -m layer1.train --mode mock --episodes 20
 
     # Evaluate a single prompt