Fix val_loss metric key ambiguity for HPO callbacks

With on_step=True and on_epoch=True, Lightning creates val_loss_step
and val_loss_epoch keys, making the plain val_loss key unreliable in
callback_metrics. This caused ModelCheckpoint, EarlyStopping, and the
HPO trial result to silently fail. Setting on_step=False ensures the
metric is logged only as val_loss at epoch end.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: forklady42

src/electrai/lightning.py

@@ -38,7 +38,12 @@ def training_step(self, batch):
     def validation_step(self, batch):
         loss = self._loss_calculation(batch)
         self.log(
-            "val_loss", loss, prog_bar=True, on_step=True, on_epoch=True, sync_dist=True
+            "val_loss",
+            loss,
+            prog_bar=True,
+            on_step=False,
+            on_epoch=True,
+            sync_dist=True,
         )
         return loss