1 epoch, more frequent evals — fits budget, catches overfitting early

- epochs: 3 -> 1 (old run showed val_loss regressing by step 2000)
- warmup_steps: 3000 -> 1000 (proportional to shorter run)
- eval_steps: 1000 -> 500 (14 evals to find generalization peak)
- save_steps: 5000 -> 2000 (more checkpoints in shorter run)

Author: brian-c-moore

configs/reasoning_core_204m.yaml

@@ -21,15 +21,15 @@ data:
 
 training:
   output_dir: ./checkpoints/reasoning_core
-  epochs: 3
+  epochs: 1
   batch_size: 1
   gradient_accumulation: 64
   lr: 2.0e-4
-  warmup_steps: 3000
+  warmup_steps: 1000
   weight_decay: 0.01
   max_grad_norm: 1.0
   fp16: true
   logging_steps: 100
-  save_steps: 5000
-  eval_steps: 1000
+  save_steps: 2000
+  eval_steps: 500
   run_name: leanformer-reasoning-core-204m