feat: child graph architecture — 5 composable sub-graphs in single parent launch
Replaces 6 separate CUDA graphs (graph_forward, graph_forward_ddqn,
graph_adam, graph_aux, graph_spectral, graph_mega) with composable
child sub-graphs assembled into a single parent graph.
Architecture:
parent_graph (single cuGraphLaunch per step)
├── spectral_child (spectral norm)
├── forward_child (cuBLAS trunk + temporal + ISV + loss + backward)
├── ddqn_child (Double DQN Pass 3)
├── aux_child (HER + EMA + IQL + IQN + attention + CQL)
└── adam_child (mamba2 bwd + pruning + grad_norm + adam + ISV update)
Each child captured independently via cudarc begin/end_capture,
composed via cudaGraphAddChildGraphNode with sequential dependencies.
Only parent is instantiated and launched.
Removed: SendSyncGraph, RawCudaGraph, capture_training_graphs,
replay_forward, replay_adam, aux_frequency, all old graph fields.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
0f8395450f