Critical: val_metrics unbound when checkpoint_frequency != validation_frequency — UnboundLocalError in training loop

[bradsmithmba]

Summary

val_metrics is only assigned inside if epoch % self.config.validation_frequency == 0. The periodic checkpoint branch references val_metrics in a ternary. When checkpoint_frequency < validation_frequency, the first checkpoint epoch will raise UnboundLocalError: local variable 'val_metrics' referenced before assignment, crashing training.

Evidence

src/training/trainer.py line 356: val_metrics = self.validate(...) — only assigned on validation epochs.
src/training/trainer.py line 398:

self.save_checkpoint(f'checkpoint_epoch_{epoch}.pth', epoch,
    val_metrics if epoch % self.config.validation_frequency == 0 else None)

val_metrics in the ternary condition is unbound on any epoch that is a checkpoint epoch but not a validation epoch.

Fix

Initialize val_metrics = None before the training loop, or inline the condition: always pass None as metrics unless this epoch also ran validation.