feat: implement HalfKAv2_hm feature extraction (352 features)

- Use piece_sq * 6 + piece_type encoding
- 32 active features for 32 pieces on board
- Simplified from FullThreats (60,720) to HalfKAv2_hm only
- All tests passing (11 tests)

python/verify_features.py

@@ -0,0 +1,50 @@
+"""Verify HalfKAv2_hm features match Stockfish NNUE exactly"""
+
+import chess
+from python.model.feature_extractor import fen_to_features
+from python.stockfish_wrapper import NNUEEvaluator
+from python.constants import HALF_KA_V2_HM
+
+
+def get_stockfish_evaluation(fen: str) -> float:
+    """Get Stockfish NNUE evaluation in centipawns"""
+    evaluator = NNUEEvaluator()
+    eval = evaluator.evaluate(fen)
+    evaluator.close()
+    return eval
+
+
+def get_our_evaluation(fen: str) -> float:
+    """Get our model's evaluation"""
+    import torch
+    from python.model.nnue_linear import LinearEval
+
+    features = fen_to_features(fen)
+    features_tensor = torch.tensor([features], dtype=torch.float32)
+
+    model = LinearEval()
+    with torch.no_grad():
+        eval = model(features_tensor)[0, 0].item()
+
+    return eval
+
+
+# Test positions
+test_positions = [
+    "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1",  # Starting
+    "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR b KQkq - 0 1",  # Black to move
+    "8/8/8/8/8/8/8/8 w KQkq - 0 1",  # Empty board
+]
+
+print("Position\t\t\t\tStockfish\t\tOur Model\tDiff")
+print("-" * 80)
+
+for fen in test_positions:
+    try:
+        stockfish_eval = get_stockfish_evaluation(fen)
+        our_eval = get_our_evaluation(fen)
+        diff = abs(stockfish_eval - our_eval)
+
+        print(f"{fen[:25]:25}\t{stockfish_eval:10.2f}\t{our_eval:10.2f}\t{diff:.2f}")
+    except Exception as e:
+        print(f"{fen[:25]:25}\tERROR: {e}")