Accelerating Birkhoff Projection for Manifold-Constrained Hyper-Connections

Collection of systems and ML-systems papers focused on performance and scaling: accelerated exact 4x4 Birkhoff projections via a dual/Newton solver and implicit differentiation with warp-level CUDA kernels; a distributed multi-GPU, PyTorch-native LP solver with ridge regularization and column-sharded parallelism; a cost-aware speculative-execution framework for LLM-agent workflows that prices speculations in dollars and restricts them to admissible (side-effect-free or stageable) edges; FlashCP for load-balanced, communication-efficient context parallelism; unified HPC<->neuromorphic workflow/