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

main.py

@@ -2,8 +2,8 @@ import argparse
 import sys
 import os
 from config import MazeConfig
-from maze import generate
-from agent import Agent, NeuralNet
+from maze import generate, bfs_distances, find_short_seed
+from agent import Agent, Genome
 from genetic import next_generation
 
 
@@ -57,10 +57,26 @@ def main():
     goal_row = cfg.maze_rows - 1
     goal_col = cfg.maze_cols - 1
 
-    grid = generate(cfg.maze_rows, cfg.maze_cols, seed=args.seed)
+    if args.seed is None:
+        print("Auto-selecting a maze with a short solution path...")
+        chosen_seed, grid, bfs = find_short_seed(cfg.maze_rows, cfg.maze_cols)
+        print(f"Using seed={chosen_seed}, BFS path={int(bfs[0,0])} steps")
+    else:
+        grid = generate(cfg.maze_rows, cfg.maze_cols, seed=args.seed)
+        bfs = bfs_distances(grid, goal_row, goal_col)
+        path_len = int(bfs[0, 0])
+        print(f"Maze BFS path: {path_len} steps")
+        if path_len > cfg.maze_rows * cfg.maze_cols // 3:
+            print(f"  [hint: long path ({path_len} steps) -- try omitting --seed for auto-selection]")
+
+    def make_agents(genomes):
+        agents = [Agent(g, 0, 0) for g in genomes]
+        for a in agents:
+            a._bfs = bfs
+        return agents
 
     # Initialize population
-    nets = [NeuralNet() for _ in range(cfg.population)]
+    nets = [Genome(cfg=cfg) for _ in range(cfg.population)]
 
     print(f"Starting training: {args.generations} generations, population={cfg.population}")
 
@@ -69,7 +85,7 @@ def main():
         print("Fast mode: rendering disabled during training.")
 
         for gen in range(args.generations):
-            agents = [Agent(net, 0, 0) for net in nets]
+            agents = make_agents(nets)
             run_generation_headless(agents, grid, goal_row, goal_col, cfg)
 
             for agent in agents:
@@ -94,7 +110,8 @@ def main():
         clock = pygame.time.Clock()
         viz = Visualizer(screen, grid, cfg)
 
-        replay_agents = [Agent(net, 0, 0) for net in nets]
+        replay_agents = make_agents(nets)
+
         ok = run_generation_visual(replay_agents, grid, goal_row, goal_col, cfg, viz, clock, args.speed, args.generations, args.steps_per_frame)
         if ok:
             print("Done. Press ESC or close window to exit.")
@@ -120,7 +137,7 @@ def main():
         viz = Visualizer(screen, grid, cfg)
 
         for gen in range(args.generations):
-            agents = [Agent(net, 0, 0) for net in nets]
+            agents = make_agents(nets)
 
             ok = run_generation_visual(agents, grid, goal_row, goal_col, cfg, viz, clock, args.speed, gen, args.steps_per_frame)
             if not ok:
@@ -140,7 +157,8 @@ def main():
 
         # Final visual replay with best generation
         print("\nTraining complete. Showing final generation replay...")
-        replay_agents = [Agent(net, 0, 0) for net in nets]
+        replay_agents = make_agents(nets)
+
         viz.reset_trails()
         run_generation_visual(replay_agents, grid, goal_row, goal_col, cfg, viz, clock, args.speed, args.generations, args.steps_per_frame)