Reinforcement Learning Runtime Entrypoint

This is the runtime umbrella for typed reinforcement-learning utilities. The modules here are executable infrastructure: they define small MDP/bandit interfaces, value-learning and policy-gradient helpers, replay buffers, checked rollout boundaries, Gymnasium subprocess sessions, and PPO rollout infrastructure.

The split is intentional:

• Runtime.RL.Core contains tensor-shaped rollout/loss helpers shared by algorithms;
• Runtime.RL.Replay contains typed bounded replay buffers for off-policy algorithms;
• Runtime.RL.Algorithms.* contains pure/mostly-pure bandit, tabular, value-learning, and policy-gradient equations;
• Runtime.RL.Boundary records host-side rollout contracts before converting observations and rewards into TorchLean tensors;
• Runtime.RL.Gymnasium is an external-process bridge and therefore a trust boundary;
• Runtime.RL.PolicyGradient.Autograd contains differentiable actor/critic losses over TorchLean refs; model architectures remain in GraphSpec/API model layers;
• Runtime.RL.PPO contains rollout/sample construction;
• Runtime.RL.Numerics.* contains optional checked float32 and interval diagnostics for RL recursions.