Changes in version 0.1.0                        

  - Added the first release of the DuckDB-native semantic graph, lexical
    grounding, dense retrieval, exact late-interaction, structured
    judgment, and benchmark APIs.
  - Validated the complete HPO 2026-06-23 release: 19,836 active
    nodes, 50,029 aliases, 24,378 direct edges, and 202,740 materialized
    is_a closure rows. Full model runs persisted 19,836 EmbeddingGemma
    vectors and 182,884 native LFM2.5-ColBERT token vectors.
  - Corrected OBO qualifier parsing, decoded escaped quoted text, and
    imported alternate identifiers. Full HPO parsing now leaves no
    dangling qualified is_a targets.
  - Made graph writes transactional and idempotent, restored graph
    indexes after projection and closure replacement, and isolated
    late-interaction blocks by provider, subject kind, subject, and
    block identifier.
  - Deduplicated lexical aliases per span and node while assigning
    unique IDs to ambiguous candidates. Structured judgments now require
    exact candidate coverage, valid confidence values, and replacements
    grounded in supplied candidates or graph context.
  - Made embedding caches content- and provider-aware, atomic,
    resumable, and safe to refresh without deleting unrelated files.
  - Registered S7 methods during namespace loading and tightened table,
    vector, token metadata, benchmark, and model-provider validation.
  - Replaced causal pooled-state retrieval paths with native
    EmbeddingGemma and LFM2.5-ColBERT providers. DuckDB VSS/HNSW remains
    an optional dense candidate-generation layer.
  - Removed the reticulate/FastHPOCR and Rfmalloc runtime surfaces.
    Ontology and local graph sources now share one graph schema and
    durable DuckDB vector storage.