New method: Tree-of-Thoughts deliberation (generate → score → prune → expand)

[hherb]

LLM-native problem-solving method exploiting parallel candidate exploration:

1. Generate — each participant proposes k distinct solution approaches (anonymised to avoid anchoring).
2. Score — all participants score all candidates on feasibility/impact/risk (reuse matrix-scoring machinery).
3. Prune — keep the top 2-3 (beam).
4. Expand — each surviving candidate gets a deep-dive round: refine, identify obstacles, concretise.
5. Repeat 2-4 until convergence or depth limit, then synthesise.

Genuinely iterative, so it depends on phase-machine loop support (#22). Together with NGT (#24) it would give the platform a real generative/exploratory arm to complement the evaluative methods.