feat: update optimization defaults for viable CMA-ES convergence

- generations: 20 -> 200 (need at least 200 for high-dim search)
- population_size: 8 -> 64 (was far below CMA-ES recommendation)
- Add pca_dimensions: 128 for dimensionality reduction
- Add use_diagonal_cma: true for sep-CMA-ES variant
- Add cache_fitness: true to avoid redundant API calls

Previous defaults (popsize=8, gen=20) gave only 160 evaluations,
which is ~1% of the theoretical minimum needed for convergence.

config.yaml

@@ -9,7 +9,11 @@ openai:
   api_key: "env" 
 
 optimization:
-  surrogate_model: "microsoft/phi-2"
+  surrogate_model: "microsoft/phi-2"  # keep for now, will be made configurable
   trigger_length: 15
-  generations: 20
-  population_size: 8
+  generations: 200        # was 20, need at least 200 for any convergence
+  population_size: 64     # was 8, need at least 64 for high-dim search
+  sigma: 0.5
+  pca_dimensions: 128     # NEW: PCA reduction target dimensions
+  use_diagonal_cma: true  # NEW: use sep-CMA-ES for high-dimensional efficiency
+  cache_fitness: true     # NEW: cache token_ids -> fitness to avoid redundant API calls