fix(path-c): reflect CUDA bridge buffer writes back to owner (fused forward was silently dropped)

run_path_c_direct_fusion_chain_route runs segment artifacts against a route-local buffers
dict. On Metal the kernel mutates caller-owned arrays in-place (zero-copy); on CUDA the
MLX<->torch bridge writes results as NEW mlx arrays into the route-local dict only, and the
route never reflected them back to logical_owner.buffers. So value_and_grad re-read the owner
(still seeded zeros) after the forward route -> fused forward outputs silently dropped ->
degenerate loss -> near-zero cotangents -> near-zero fused grads. This made fused-CUDA
training silently produce GARBAGE gradients.

Fix: after the segment loop, reflect the route-local buffers back into logical_owner.buffers /
logical_buffers. CUDA-gated (_path_c_default_target()=="cuda") so Metal is byte-for-byte
unchanged; env CPPMEGA_PATH_C_DISABLE_CUDA_WRITEBACK=1 disables it for verification calibration.

Found + verified by a standalone b0 ABI verifier (drives the real fused mamba3 backward via the
production banked-ABI runtime at seq=8): with the fix the fused forward delta matches eager to
~2% and loss to 0.002 (was: total changed 0/71, output exactly 0.0). Metal direct-chain route
tests 7 passed.

Author: jeweldave

scripts/m04_train_step.py

@@ -6419,6 +6419,30 @@ def run_path_c_direct_fusion_chain_route(
                 "result_type": type(result).__name__,
             }
         )
+    # CUDA-only write-back fix: the MLX<->torch bridge replaces caller-owned bank
+    # arrays with NEW mlx arrays holding the kernel results, living only in the
+    # route-local ``buffers`` dict. Metal mutates the owner's arrays in-place
+    # (zero-copy) and needs none of this. The caller
+    # (PathCDirectFusionChainTrainingRuntime.value_and_grad) re-reads
+    # ``logical_owner.buffers`` after the forward route to seed the loss bridge
+    # and backward, so without reflecting the writes back the fused forward
+    # outputs are silently dropped -> degenerate loss -> near-zero fused grads.
+    # CUDA-gated so Metal is byte-for-byte unchanged.
+    # (env CPPMEGA_PATH_C_DISABLE_CUDA_WRITEBACK=1 disables it to reproduce the
+    #  pre-fix broken state for verification calibration.)
+    if (
+        _path_c_default_target() == "cuda"
+        and os.environ.get("CPPMEGA_PATH_C_DISABLE_CUDA_WRITEBACK", "") != "1"
+    ):
+        for _target_map in (getattr(logical_owner, "buffers", None), logical_buffers):
+            if _target_map is None:
+                continue
+            try:
+                for _name, _value in buffers.items():
+                    _target_map[_name] = _value
+            except (TypeError, AttributeError):
+                # immutable mapping (e.g. MappingProxyType): cannot write back.
+                pass
     return {
         "status": "ok",
         "runtime_uses_direct_fusion_chain": True,