Align GRPOConfig defaults with CLI: 10 steps, 7 episodes

The dataclass defaults were still 50 steps / 10 episodes, causing
deployments that don't pass explicit CLI args to run much longer
than intended.

https://claude.ai/code/session_01DPirJ78YYN4fJUvUFJ5D6V

layer1/grpo_trainer.py

@@ -37,8 +37,8 @@ class GRPOConfig:
 
     # GRPO
     num_candidates: int = 4         # N candidate prompts per step
-    episodes_per_candidate: int = 10  # K episodes to evaluate each candidate
-    num_training_steps: int = 50
+    episodes_per_candidate: int = 7   # K episodes to evaluate each candidate
+    num_training_steps: int = 10
     learning_rate: float = 5e-5
     max_prompt_length: int = 512
 

layer1/train.py

@@ -3,7 +3,7 @@ Layer 1 — Executable GRPO training script.
 
 Usage:
     # Full GPU training (requires Colab/GPU + train deps)
-    python -m layer1.train --mode train --steps 50
+    python -m layer1.train --mode train --steps 10
 
     # CPU mock optimization (evaluates hand-written prompts)
     python -m layer1.train --mode mock --episodes 20