Add architecture documentation explaining convergence algorithm and scoring

[that-github-user]

Summary

The convergence algorithm (Jaccard similarity + union-find clustering) and recommendation scoring are non-trivial but only documented in code comments. Contributors and users need to understand:

1. How convergence works — diff parsing → Jaccard similarity → single-linkage clustering
2. How scoring works — test pass (100pts) + convergence (0-50pts) + diff size (0-10pts)
3. Why these choices — why Jaccard over cosine? why 0.3 threshold? why 50/50 weight?
4. Limitations — what the algorithm misses (semantic equivalence, rename-only changes)

Proposed

Create docs/architecture.md with:

• System diagram (prompt → parallel agents → convergence → recommendation)
• Convergence algorithm explanation with examples
• Scoring formula with rationale
• Known limitations and future improvements

This is also valuable content for the README and for academic citations.

[that-github-user]

Fixed in #48. Added docs/architecture.md with system overview diagram, module responsibilities, convergence algorithm details (Jaccard similarity + union-find clustering), recommendation scoring formula with weight rationale, security model, and known limitations.