feat: add hypo_critic and hypo_rewrite in proposal

[Hoder-zyf]

Description

This PR introduces an automated “Critique → Rewrite” stage into the data‑science proposal flow:

• Prompt additions

  • Adds hypothesis_critique and hypothesis_rewrite blocks in prompts_v2.yaml, complete with JSON schemas and guidance to avoid over‑specification and historical overfit.

• Pipeline implementation

  • hypothesis_critique() – scores each generated hypothesis on feasibility, alignment, and improvement direction, returning constructive feedback.
  • hypothesis_rewrite() – produces a single, decisive, testable hypothesis per item, integrating critique insights while preserving innovation.
  • Robust validation & fallback logic ensure all hypotheses receive usable critiques even if the model output is partially malformed.

Motivation and Context

Raw, auto‑generated hypotheses are often vague or at high risk.
By forcing a critique → guided rewrite loop we expect:

1. Higher alignment with real challenge pain‑points.
2. More actionable, test‑ready experiments.
3. Better balance between safe incremental gains and bold innovation.

This should cut wasted GPU hours on un‑implementable ideas and boost leaderboard progress.

How Has This Been Tested?

• If you are adding a new feature, test on your own test scripts.

Screenshots of Test Results (if appropriate):

1. Your own tests:

Types of changes

• Fix bugs
• Add new feature
• Update documentation

---

📚 Documentation preview 📚: https://RDAgent--1106.org.readthedocs.build/en/1106/