semantic-pkm-creatorlisted

# Scene–Concept–Entity Extraction Input is a batch of linear source texts. Output `sce-scenes.yaml`, `sce-concepts.yaml`, and `sce-entities.yaml` in the user-selected directory, with progress in `_sce-round1-progress.md`. The value is reuse: one scene can directly invoke a concept established elsewhere. ## Principles 1. **Source first** — `define` and IPO come from the source and stay within 50 words. Leave unsupported content empty rather than inventing it. 2. **Concepts are the hub** — Establish concepts first, attach entities to concepts, and let scenes invoke only established concepts and entities. 3. **Prefer omission** — A candidate without its own IPO or decomposition is not a concept. Merge abstractions into their parent and demote fragments to steps. 4. **Bound references** — Keep two to six relations per concept and only the strongest relation for a knowledge-element pair. 5. **Stop for confirmation between rounds.** Never start round two before approval. ## Round One: Skeleton Process five to ten texts per batch and update `_sce-round1-progress.md`. 1. Read the complete batch and record progress. 2. Extract candidates with only `id + define + sources`: scenes answer “how do I…”, concepts can express an IPO, and entities answer “did I use X?” 3. Deduplicate globally: merge at 80% or greater semantic overlap, normalize synonymous IDs, and return overly fine steps to their parent. Record every `X → Y` merge. 4. Write the three YAML files and ma