fix: genome GA converges via BFS fitness + two-rate prefix-aware mutation

- Replace Manhattan distance with BFS distances (actual maze path length)
- Switch from NN to direct sequence genome (what YouTube Shorts actually use)
- Mutation-only GA (crossover breaks positional maze paths)
- Two-rate mutation: low rate before best-step (preserve prefix), high after (explore tail)
- Auto-seed selection finds maze with short BFS path
- Default maze 15x15, cell_size=36, max_steps=150
- Typically converges in 5-15 generations

config.py

@@ -3,21 +3,23 @@ from dataclasses import dataclass
 @dataclass
 class MazeConfig:
     # Maze geometry
-    maze_cols: int = 25
-    maze_rows: int = 25
-    cell_size: int = 24
+    maze_cols: int = 15
+    maze_rows: int = 15
+    cell_size: int = 36
     wall_width: int = 2
 
     # GA / Training
     population: int = 100
-    max_steps: int = 300
+    max_steps: int = 150
     elite_frac: float = 0.30
-    mutation_std: float = 0.05
-    mutation_rate: float = 0.20
+    n_elite: int = 3          # few elites → maintains population diversity
+    mutation_std: float = 0.08
+    mutation_rate: float = 0.35
+    temperature: float = 0.7  # softmax sampling temperature (lower = more greedy)
 
-    # Neural net
-    n_inputs: int = 8
-    n_hidden: int = 12
+    # Neural net — 14 inputs: 4 walls + 2 goal dir + 2 pos + 4 last-action + 1 blocked + 1 revisit
+    n_inputs: int = 14
+    n_hidden: int = 16
     n_outputs: int = 4
 
     # Fitness