- Move torch/transformers to [nli] optional deps (base install ~1MB)
- CoherenceScorer auto-detects NLI when torch+transformers available
- Scorer delegates to NLIScorer for both logical and factual divergence
- NLIScorer accepts configurable model_name for fine-tuned models
- retrieve_context returns values only (better NLI premises)
- CoherenceAgent disables NLI in mock mode (heuristic-designed data)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>