Phase 4: training loop (code only — for user to run)

- src/train.py: AdamW (decay on matrices only), cosine LR w/ warmup, grad
  clipping, periodic train/val loss, CSV logging, resumable checkpoints.
  --overfit mode runs the overfit-one-batch sanity check. MPS, fp32.
- configs/v1.py: TrainConfig (batch 64, peak LR 6e-4, 20k iters, etc.)
- tests/test_train.py: LR schedule + optimizer grouping (3 tests, no run)

configs/v1.py

@@ -21,5 +21,30 @@ class ModelConfig:
     d_ff: int = 1024             # MLP hidden width (4 x d_model)
 
 
-# The single source of truth other modules import.
+@dataclass(frozen=True)
+class TrainConfig:
+    # Data / batching. tokens per step = batch_size * context_length = 16,384.
+    batch_size: int = 64
+
+    # AdamW optimiser.
+    learning_rate: float = 6e-4  # peak LR (reached after warmup)
+    min_lr: float = 6e-5         # cosine-decay floor (~10% of peak)
+    warmup_iters: int = 200
+    max_iters: int = 20_000      # ~327M tokens (~0.6 epoch of the 555M corpus)
+    weight_decay: float = 0.1    # applied to matrices only, not norms/biases
+    beta1: float = 0.9
+    beta2: float = 0.95
+    grad_clip: float = 1.0       # clip global grad norm to this
+
+    # Evaluation, logging, checkpointing (all in optimiser steps).
+    eval_interval: int = 500     # how often to measure train/val loss
+    eval_iters: int = 100        # batches averaged per loss estimate
+    log_interval: int = 20       # how often to print step/loss/lr/speed
+    checkpoint_interval: int = 1000
+
+    seed: int = 1337
+
+
+# Single source of truth other modules import.
 model = ModelConfig()
+train = TrainConfig()