feat: Test Time RL - Auto Research

[Aum08Desai]

Summary

• adds a thin TTT-Discover integration on top of autoresearch
• keeps the original autoresearch workflow intact
• uses the outer RL loop to propose train.py edits and uses inner-loop val_bpb as reward
• This should hypothetically (based on literature), massively outperform any frontier model, even when using TTT-Discover on a small OS model.

Notes

• this has now been mostly tested and is confirmed to work in large part end to end
• I am currently still debugging it locally, but it works almost in entirety
• the main remaining rough edge is that the agent's code-edit replacement behavior can still be somewhat unreliable in some batches
• the most recent pushes are specifically aimed at improving that edit reliability

Safety

• this PR does not overwrite upstream master
• it is proposed from a separate fork branch for review