feat(v7-a01): llama3_8b preset e2e at scaled hidden=512

Closes V7-A01 (cppmega-mlx-c30.1.5): per user redirect
('конструктор из кирпичиков'), pins that the llama3_8b preset's
brick composition runs through stage_train at hidden=512 with
finite losses + non-zero weight delta. Full H=4096 perf gate
lives under V7-I05 (production-preset baseline).

Tests (tests/v4/test_llama3_8b_scaled_e2e.py): 2/2 —
  * llama3_8b preset at H=512 runs through train(2) with finite
    losses + weight_delta > 0.
  * Preset composition exposes attention + mlp brick kinds.

Author: apstenku123

tests/v4/test_llama3_8b_scaled_e2e.py

@@ -0,0 +1,70 @@
+"""V7-A01: llama3_8b preset instantiated end-to-end at scaled dims.
+
+Per user redirect ("у нас же конструктор из кирпичиков"), the
+acceptance is that the llama3_8b preset's brick composition runs
+through stage_train at a realistic-but-mini scale (H=512 with
+multi-repeat). Hidden=4096 full-scale is HBM-bound on a single
+Mac and tracked as a separate perf gate (V7-I05).
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from cppmega_v4.architectures.presets import build_preset_specs
+from cppmega_v4.jsonrpc.schema import VerifyParams
+from cppmega_v4.runner import Pipeline, run_pipeline
+
+
+def _spec_from_preset(name: str, hidden: int = 512) -> VerifyParams:
+    specs = build_preset_specs(name, hidden_size=hidden)
+    return VerifyParams.model_validate({
+        "graph": {
+            "nodes": [
+                {"id": f"n{i}", "kind": s["kind"],
+                 "params": s.get("params", {})}
+                for i, s in enumerate(specs)
+            ],
+            "edges": [
+                {"src": f"n{i}", "dst": f"n{i + 1}"}
+                for i in range(len(specs) - 1)
+            ],
+        },
+        "dim_env": {"B": 1, "S": 8, "H": hidden,
+                    "nh": max(2, hidden // 64),
+                    "nkv": max(1, hidden // 128),
+                    "head_dim": 64,
+                    "num_experts": 4, "top_k": 2},
+        "loss": {"kind": "cross_entropy",
+                 "head_outputs": [f"n{len(specs) - 1}"]},
+        "optim": {"kind": "adamw",
+                  "groups": [{"matcher": "all", "lr": 1e-3,
+                              "weight_decay": 0.01,
+                              "betas": [0.9, 0.95]}]},
+    })
+
+
+def test_v7_a01_llama3_8b_preset_runs_at_h512():
+    rep = run_pipeline(
+        _spec_from_preset("llama3_8b", hidden=512),
+        Pipeline.from_dict({
+            "stages": ["parse", "verify_build_spec", "build_model",
+                       "train"],
+            "stage_options": {"train": {"num_steps": 2}},
+        }),
+    )
+    tr = next(s for s in rep.stages if s.name == "train")
+    assert tr.status == "ok", f"train failed: {tr.error}"
+    assert len(tr.extras["losses"]) == 2
+    for L in tr.extras["losses"]:
+        assert -1e10 < L < 1e10
+    assert tr.extras["weight_delta_norm"] > 0
+
+
+def test_v7_a01_preset_compose_yields_known_brick_kinds():
+    """llama3_8b preset must produce attention + mlp bricks (the
+    canonical LLaMA repeat unit)."""
+    specs = build_preset_specs("llama3_8b", hidden_size=128)
+    kinds = {s["kind"] for s in specs}
+    assert "attention" in kinds
+    assert "mlp" in kinds