Add SIA: Self-Improving AI with Harness & Weight Updates

[imnida]

Implements the SIA framework from Hebbar et al. (arXiv:2605.27276).
The loop lets a Feedback-Agent iteratively improve both the scaffold
(harness) and the model weights (LoRA) of a task-specific agent.

Key components:

• sia_loop.py — main configurable loop (Meta-Agent → execute → Feedback-Agent)
• meta_agent.py — generates initial scaffold A1 using Claude Sonnet 4.6
• feedback_agent.py — analyses trajectory τg, decides harness vs weight update
• task_agent.py — executes scaffold against dataset, captures trajectory
• trajectory.py — structured execution log (Step, ToolCall, Trajectory)
• verifier.py — deterministic per-instance reward interface
• weight_updates/ — six RL algorithms: PPO+GAE, GRPO, Entropic Advantage
  Weighting, REINFORCE+KL, Best-of-N BC, DPO
• tasks/ — three benchmark tasks: LawBench (191-class Chinese legal),
  AlphaEvolve TriMul (CUDA kernel), MAGIC scRNA-seq denoising

https://claude.ai/code/session_01DLqnGSQGNhPHnUzTLgJ6id