v1 scope: the AI-ready-data spine, nothing else

Framework v1 ships exactly the Stages that tell the "messy data → AI-ready data" story end to end. Doc level: Parse (document → markdown + page map), Chunk, Entity Extraction, Relation Extraction, PII Redaction. Corpus level: Profiling (statistical/quality features), Disambiguation (entity linking + cross-document resolution), Graph Assembly (canonical KG with per-edge Evidence provenance), Export (Parquet/JSONL/TTL/GraphML). Explicitly OUT of core, available later as plugins: text compression, embedding/retrieval serving (ColBERT/ColPali), trace/groundedness scoring, link prediction, and formal ontology induction (SHACL/OWL). These are deferred for sequencing and dependency-surface reasons, NOT because the code is weak: the maintainer's own review of the mined algorithmic assets rates the ontology-induction stack (ConceptInducer/Taxonomy/AxiomInducer) at 4–4.5/5 with HIGH reinvention cost, and the link-prediction stack includes an inductive/zero-shot ULTRA scorer (ICLR 2024), not merely transductive RotatE. They are preserved verbatim as scheduled post-v1 plugins (P1–P4), not discarded. Rationale: each Stage costs a Capability protocol, a reference Provider, and documentation maintained by one author; a sharp spine beats broad parity, and the corpus-level KG is the differentiator versus parse-only tools like docling/unstructured, so doc-level-only was rejected.