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/20261028-180840_spike-qkon-lr1e-3/config.json

@@ -0,0 +1,38 @@
+{
+  "out_dir": "/Users/nik/ISE/9.CS6513/thesis/06-experiments/runs/20261028-180840_spike-qkon-lr1e-3",
+  "eval_interval": 100,
+  "log_interval": 10,
+  "eval_iters": 20,
+  "eval_only": false,
+  "always_save_checkpoint": false,
+  "init_from": "scratch",
+  "wandb_log": false,
+  "wandb_project": "owt",
+  "wandb_run_name": "gpt2",
+  "dataset": "tinystories_spike",
+  "gradient_accumulation_steps": 1,
+  "batch_size": 32,
+  "block_size": 256,
+  "n_layer": 4,
+  "n_head": 4,
+  "n_embd": 256,
+  "dropout": 0.0,
+  "bias": false,
+  "qk_norm": true,
+  "seed": 1337,
+  "data_root": "/Users/nik/ISE/9.CS6513/thesis/06-experiments/data",
+  "learning_rate": 0.001,
+  "max_iters": 300,
+  "weight_decay": 0.1,
+  "beta1": 0.9,
+  "beta2": 0.95,
+  "grad_clip": 1.0,
+  "decay_lr": true,
+  "warmup_iters": 30,
+  "lr_decay_iters": 300,
+  "min_lr": 0.0001,
+  "backend": "nccl",
+  "device": "mps",
+  "dtype": "float32",
+  "compile": false
+}