feat(v5-g25): MoE/sparse_moe/bailing_moe bricks → extras.moe observability

Closes V5-G25 / cppmega-mlx-zuf. stage_train detects MoE-family brick
in wire graph; extras.moe = {kind, num_experts, top_k,
routing_entropy=null, load_balance_loss=null, dropped_token_ratio=null}.
Real routing entropy / load-balance loss / dropped-token measurement
deferred to v6+ (requires expert-routing forward hook).

2/2 pytest green.

Author: apstenku123

cppmega_v4/runner/stages.py

@@ -692,6 +692,28 @@ def _count(tree: Any) -> int:
             "num_clips": 0,
         }
 
+        # G25: detect MoE / sparse-experts bricks in graph; surface
+        # routing config in extras.moe. Pure observation — actual
+        # routing/load-balance computation is v6+ work.
+        moe_extras: dict[str, Any] | None = None
+        try:
+            moe_kinds = {"moe", "bailing_moe", "sparse_moe"}
+            wire_nodes = getattr(ctx.spec.graph, "nodes", []) or []
+            moe_node = next((n for n in wire_nodes
+                             if getattr(n, "kind", "") in moe_kinds), None)
+            if moe_node is not None:
+                p = dict(getattr(moe_node, "params", {}) or {})
+                moe_extras = {
+                    "kind": str(getattr(moe_node, "kind", "moe")),
+                    "num_experts": int(p.get("num_experts", 1)),
+                    "top_k": int(p.get("top_k", 1)),
+                    "routing_entropy": None,        # v6: actual measurement
+                    "load_balance_loss": None,
+                    "dropped_token_ratio": None,
+                }
+        except Exception:
+            pass
+
         # G07: read precision toggles from spec (passthrough — backend
         # doesn't actually switch dtype yet, but extras report what the
         # UI asked for so e2e can assert propagation). Real mixed/fp8
@@ -921,6 +943,7 @@ def _count(tree: Any) -> int:
                 "train_dtype": train_dtype,
                 "master_dtype": master_dtype,
                 "fp8_active": fp8_active,
+                "moe": moe_extras,
                 "opt_state_carried": opt_state_carried,
                 "run_id": run_id,
                 "mtp": _compute_mtp_extras(

tests/v5/test_stage_train_moe.py

@@ -0,0 +1,58 @@
+"""G25: MoE bricks detected in graph → extras.moe populated."""
+
+from __future__ import annotations
+
+from cppmega_v4.jsonrpc.schema import VerifyParams
+from cppmega_v4.runner import Pipeline, run_pipeline
+
+
+def _spec(nodes: list[dict]) -> VerifyParams:
+    return VerifyParams.model_validate({
+        "graph": {
+            "nodes": nodes,
+            "edges": [{"src": nodes[i]["id"], "dst": nodes[i+1]["id"]}
+                      for i in range(len(nodes) - 1)],
+        },
+        "dim_env": {"B": 1, "S": 8, "H": 32, "nh": 2, "nkv": 1, "head_dim": 16},
+        "loss": {"kind": "cross_entropy",
+                 "head_outputs": [nodes[-1]["id"]]},
+        "optim": {"kind": "adamw",
+                  "groups": [{"matcher": "all", "lr": 1e-3,
+                              "weight_decay": 0.01, "betas": [0.9, 0.95]}]},
+    })
+
+
+def _run(nodes: list[dict]) -> dict:
+    # Skip build_model — MoE brick has V4MoEConfig kwargs that build_model
+    # rejects; G25 is observation-only (extras populated from raw graph).
+    report = run_pipeline(_spec(nodes), Pipeline.from_dict({
+        "stages": ["parse", "verify_build_spec", "train"],
+        "stage_options": {"train": {"num_steps": 2}},
+    }))
+    train = next(s for s in report.stages if s.name == "train")
+    return train.extras
+
+
+def test_no_moe_brick_extras_moe_is_none():
+    extras = _run([
+        {"id": "attn", "kind": "attention", "params": {}},
+        {"id": "mlp", "kind": "mlp",
+         "params": {"intermediate_size": 64, "activation": "swiglu"}},
+    ])
+    assert extras.get("moe") is None
+
+
+def test_moe_brick_populates_num_experts_top_k():
+    # Use mlp (instantiates cleanly) but inject a "moe" pseudo-node so
+    # the detection logic in stage_train sees it via the wire graph.
+    extras = _run([
+        {"id": "moe", "kind": "moe",
+         "params": {"num_experts": 8, "top_k": 2}},
+        {"id": "mlp", "kind": "mlp",
+         "params": {"intermediate_size": 64, "activation": "swiglu"}},
+    ])
+    moe = extras.get("moe")
+    assert moe is not None
+    assert moe["kind"] == "moe"
+    assert moe["num_experts"] == 8
+    assert moe["top_k"] == 2