Add meta-research loop to program.md

[PodJamz]

Summary

Addresses #314. Adds a ## Meta-Research section to program.md that instructs the agent to periodically review its own experiment history and refine its research strategy.

What this adds

Every 20 experiments, the agent:

1. Scans results.tsv for repeated failure patterns (same hypothesis type discarded 3+ times)
2. Appends confirmed dead-ends to hypotheses_blacklist.md (untracked, like results.tsv)
3. Re-evaluates which research phases are producing gains vs stalling
4. Adjusts its active research strategy accordingly

What this does NOT change

• No Python files modified
• The fixed outer evaluation loop (val_bpb, 5-minute budget) is never touched
• The agent can only modify its research strategy, never its evaluation criteria

Why

The current loop optimizes train.py but the research strategy in program.md is static. After 50+ experiments, the agent may keep trying hypothesis classes that have been proven unproductive. A meta-research step lets the agent learn from its own experiment history and focus on productive directions.

This is inspired by the HyperAgent framework (Zhang, Lehman, Clune - Meta FAIR, 2025) - the principle that the improvement mechanism itself should be subject to improvement. The outer eval loop stays fixed; the research strategy evolves.