fix(path-c): seed direct-chain first brick with raw residual, not pre-normed (double-norm)

make_path_c_direct_chain_pre_step_runtime_owner seeded {first_brick}_hidden with
layers[start].norm(prefix_hidden), but the fused chain's entry-RMSNorm re-applies the
norm -> double-norm. The eager layer contract (hybrid_lm.py:687-697) is residual = hidden
(raw), then delta = route_delta(hidden) which norms internally; the chain replays that, so
the seed must be the RAW residual stream.

Fix: boundary_hidden = prefix_hidden (raw) instead of layers[start].norm(prefix).

Metal (chain start_layer=10): direct-chain runtime loss-vs-eager err 1.02e-2 -> 0.0 (exact),
grad err 1.23e-2 -> 0.0. Flag-independent, target-agnostic (corrects CUDA equally).
Tests 58 passed (direct-chain/pre-step-owner/value_and_grad/bridge). Found by the moe-span
executor investigation; extracted minimal.

Author: jeweldave

scripts/m04_train_step.py

@@ -7626,10 +7626,13 @@ def make_path_c_direct_chain_pre_step_runtime_owner(
         batch,
         end_layer_index=start_layer_index,
     )
-    layers = tuple(getattr(model, "layers", ()))
+    # The fused chain's first brick reads ``{first_brick}_hidden`` as the RAW
+    # residual baseline and applies its own entry-RMSNorm (a fused
+    # ``entry_rmsnorm`` surface bound to ``layers.{start}.norm.weight``) to
+    # derive the route input. Seeding the raw residual stream therefore matches
+    # the eager ``updated = hidden + route_delta(norm(hidden))`` contract;
+    # seeding the pre-normed hidden would double-apply the entry norm.
     boundary_hidden = prefix_hidden
-    if start_layer_index < len(layers):
-        boundary_hidden = layers[start_layer_index].norm(prefix_hidden)
     hidden_seed_names = {
         "hidden",
         f"{first_brick_name}_hidden",