Phase 3: decoder-only transformer (~4.3M params)

- src/model.py: GPT written module by module — RMSNorm, explicit causal
  self-attention (scores/mask/softmax/weighted-sum), GELU MLP, pre-norm
  residual blocks, learned positions, weight-tied head. GPT-2 style init.
- configs/v1.py: frozen ModelConfig dataclass (4L/256d/4h/256ctx/4096vocab)
- tests/test_model.py: exact param count (4,262,144), forward shapes, init
  loss ~= ln(vocab), weight tying, causal-masking check (5 tests, forward-only)
- Verified forward pass runs on MPS.

configs/v1.py

@@ -0,0 +1,25 @@
+"""Experiment v1 configuration.
+
+One experiment = one config file. Iteration 2 is a copy of this file with
+different numbers, not a code change.
+
+v1 is deliberately small (~4.3M parameters): the goal is a complete, legible
+pipeline, not a strong model. It sits below TinyStories' coherence sweet spot
+(~30M params), so some incoherence in the output is expected.
+"""
+
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class ModelConfig:
+    vocab_size: int = 4096       # matches the trained tokenizer
+    context_length: int = 256    # tokens of history the model attends over
+    n_layers: int = 4
+    n_heads: int = 4
+    d_model: int = 256           # residual-stream width
+    d_ff: int = 1024             # MLP hidden width (4 x d_model)
+
+
+# The single source of truth other modules import.
+model = ModelConfig()