Narrative scoring and thread management for fabula.

Provides the GM's evaluation toolkit — scoring signals that tell the MCTS evaluation function whether a candidate action improves the narrative.

Research foundation

• Nelson & Mateas (2005) "Search-Based Drama Management" (AIIDE 2005) — GM as optimizer with quality function over narrative states. This crate IS that quality function. The [scorer] module implements the composite evaluation from signals to score.
• Schulz et al. (2024) "Narrative Information Theory" (arXiv:2411.12907) — Five measures (Complexity, Pivot, Predictability, Suspense, Plot Twist). [pivot] implements the Pivot measure via Jensen-Shannon Divergence.
• Booth (2009) "The AI Director: Left 4 Dead's Approach to Procedural Intensity" — Pacing via trajectory sampling. [tension] classifies numeric trajectories (Rising/Falling/Plateau/Peak/Valley).
• Ely, Frankel, Kamenica (2015) "Suspense and Surprise" — Trajectory matters more than absolute value; informs tension scoring.
• Kowal, Mary Robinette. MICE Quotient (Writing Excuses) — Narrative threads open and close; well-formed stories close in reverse order (FILO). [thread] validates this nesting property.

Modules

• [thread] — Thread lifecycle management (MICE-style open/close tracking, FILO validation)
• [tension] — Numeric trajectory sampling (is tension rising/falling/plateaued?)
• [pivot] — Event distribution shift detection (Schulz 2024 JSD pivot measure)
• [scorer] — Composite narrative quality score for MCTS evaluation

Design principles

1. Scoring, not patterns. The GM needs scoring signals, not more patterns. The engine finds matches; this crate evaluates narrative quality.
2. DataSource-agnostic where possible. ThreadTracker and PivotDetector work on engine output (TickDelta, SiftEvent). TensionTracker accepts caller-provided samples, not DataSource queries.
3. Research-backed. Each module cites its academic foundation.