LLM-judge ranker

Order N artifacts by quality on a stated dimension. The comparator is "the LLM judges A vs B against this goal," done pairwise.

Project contract

Output dir: <project>/data/sst-llm-judge-ranker/<run-id>/ where <run-id> is either user-supplied or auto-generated as <utc>_<short-uuid>. Inside:
- metadata.jsonl — append-only log of every artifact registered.
- ranking_state/<artifact-set>.ranklist.json — the current ranked list (just IDs, ordered best→worst).
Tools required: harness's Read (for candidate files). No web access; ranking is pure analysis.
Input shapes:
- Direct compare: two file paths + an optional goal → returns winner + rationale.
- Batch rank: a directory of candidate files + a goal → produces an initial ranklist.
- Incremental insert: a new candidate path + the existing ranklist file → places the new candidate at its correct position with binary-search comparisons.