Tech spike: four smoke-test runs and a divergence-stop test

Pipeline works on MPS (~105 s per 300-step run), results are reproducible,
and the divergence stop works. Not experimental evidence.

Claude-Session: https://claude.ai/code/session_01S8KN7NuUCm8nXLmy6eDuVv

06-experiments/runs/20261028-181408_test-divergence-stop/metrics.jsonl

@@ -0,0 +1,2 @@
+{"type": "eval", "iter": 0, "train_loss": 10.9141, "val_loss": 10.9167, "t": 5.48}
+{"type": "train", "iter": 0, "loss": 10.9193, "ms": 4405.09, "t": 5.8}