feat: ISMCTS q-value as training label

C#: TrainingCollector新增q_value列, IsmctsBot传入winRate
CSV: reward→q_value, ISMCTS状态带真实评估值, Neural填0
Python: 优先取ISMCTS状态+评估值训练, 无q_value的Neural状态跳过
旧CSV格式(reward)自动fallback到Neural-only过滤

信号从'这局谁赢'改为'ISMCTS评估这步有多好'

scripts/training_orchestrator.py

@@ -51,14 +51,18 @@ def selfplay(games):
     return csv
 
 def load_csv_states(csv_dir):
-    """Load states from CSV. If player column present, only keep Neural (player=2)."""
+    """Load states from CSV.
+    New format (q_value): prefer ISMCTS states (player!=2) with ISMCTS evaluation as target.
+    Old format (reward): fall back to Neural-only with game result."""
     import numpy as np
     states, results = [], []
-    neural_only = 0
+    filtered = 0
+    ismcts_samples = 0
     for fp in sorted(glob.glob(f'{csv_dir}/game_*.csv')):
         try:
             with open(fp) as f:
                 header = f.readline().strip()
+                has_qvalue = 'q_value' in header
                 has_player = header.endswith(',player')
                 for line in f:
                     if line.startswith('#'): continue
@@ -67,17 +71,29 @@ def load_csv_states(csv_dir):
                     sf = [float(x) for x in p[1].split(',')]
                     rest = p[2].strip(',').split(',')
                     if len(sf) != 364: continue
-                    if has_player and len(rest) >= 3:
-                        player = int(rest[2])
-                        if player != 2:  # only train on Neural states
-                            neural_only += 1
-                            continue
                     if len(rest) < 2: continue
+                    label = float(rest[1])  # q_value or reward
+                    if has_qvalue and has_player and len(rest) >= 3:
+                        player = int(rest[2])
+                        if player == 2 and abs(label) < 0.01:
+                            # Neural state with no q_value (0.0) → skip
+                            filtered += 1
+                            continue
+                        if player != 2:
+                            ismcts_samples += 1
+                    elif not has_qvalue and has_player and len(rest) >= 3:
+                        # Old format: Neural-only filter
+                        player = int(rest[2])
+                        if player != 2:
+                            filtered += 1
+                            continue
                     states.append(sf)
-                    results.append(float(rest[1]))
+                    results.append(label)
         except: pass
-    if neural_only:
-        log(f"  filtered out {neural_only} non-Neural states")
+    if ismcts_samples:
+        log(f"  {ismcts_samples} ISMCTS q-value samples")
+    if filtered:
+        log(f"  filtered {filtered} unusable states")
     return np.array(states, dtype=np.float32) if states else None, np.array(results, dtype=np.float32) if results else None
 
 def train_v(prev_winrate=None):