liminis-context-graph v0.16.2 — assert-only ontology types (extract: false)

[verveguy]

v0.16.2 adds one community-requested ontology feature: assert-only types. It's additive, with no migration and no schema change, so there's nothing to do when upgrading from 0.16.x. An ontology that doesn't use the new flag behaves exactly as before. If you're coming from 0.15.x or earlier, read the 0.16.0 notes first.

extract: false (#636)

If your host composes a base ontology into every group, it probably includes structural types it asserts itself from metadata, not from prose. Examples are a provenance Source with DERIVED_FROM, the document entity type, or catalog kinds. Declaring those types used to offer them to the LLM extractor too, so it started creating them from prose. With identity: true, those fuzzy extractions landed in an identity-bearing kind and collided with the host's own keyed assertions.

Now you can mark such a type assert-only:

entity_types:
  - name: Source
    identity: true
    extract: false
relation_types:
  - name: DERIVED_FROM
    source_type: Document
    target_type: Source
    extract: false

An extract: false type:

• is left out of the extraction prompts;
• is never accepted from the extractor. A stray extracted label is reclassified to Unclassified, with the original label kept, in both strict and open mode. It never gets the type's identity kind.
• stays fully declared for knowledge_assert_entity / knowledge_assert_relationship, the identity set and its -32003 guard, ancestry, and drift;
• is never a retyping target for reprocessing or relation canonicalization.

The flag defaults to true, so ontologies without it produce identical prompts and hashes and upgrading causes no drift. Flipping it on a type shows as drift, and knowledge_reload_ontology (0.16.1) accepts the change without a restart. See docs/ontology.md.

What was verified

• CI: the full Rust suite and real-corpus end-to-end jobs on Linux, the Windows named-pipe job with real index creation and queries, and release builds for all four targets with the pinned-extension check.
• Feature tests: a dedicated extract_false suite covers:
  • prompt exclusion;
  • an extracted Source never getting the Source kind;
  • asserts still working;
  • no retyping;
  • byte-identical prompts and hashes without the flag;
  • flipping the flag reporting drift and a reload accepting it.
• macOS: the sidecar-gated replay battery ran with no skips.
• Published artifacts: all four archives pass their checksums, every bundled extension matches its pinned hash, and the published binary reports 0.16.2.

Full detail: docs/releases/0.16.2.md · CHANGELOG