v1.1.0 — keyless semantic search, direction-aware graphs, dynamic schema (phase 1)

Fixed

• Relation endpoint recovery — graph edges now survive small-model
  extraction. Small local models frequently emit a correct relation while
  omitting one endpoint from entities; the worker's anti-hallucination
  guard (relations never create entities) then silently dropped the edge.
  Measured on the gold fixture with llama3.2:3b: 0% of emitted relations
  had both endpoints listed — i.e. zero graph edges survived on a fresh
  keyless install. extract() now runs a best-effort second pass that asks
  the model to classify exactly the missing endpoint names against the same
  text (example-anchored prompt; anything beyond the requested names is
  discarded). Edge survival: 0% → 79%, direction accuracy unchanged at
  93%; the remainder are junk phrase-objects that are correctly rejected.
  The direction check now also measures and gates endpoint completeness
  (BRAIN_ENDPOINT_MIN_COMPLETENESS, default 0.75).

Changed

• Relationship extraction is now direction-aware. The extraction prompt
  states that relations are directed (subject → predicate → object), defines
  subject vs. object, and gives worked directional examples (including the
  passive-voice trap that flips small models). On the gold fixture this lifts
  llama3.2:3b direction accuracy from ~79% to ~93%. The prompt and provider
  calls moved to extraction.ts (importable without starting the worker loop).

Added

• Strict curation mode. BRAIN_REQUIRE_HUMAN_REVIEW=1 disables all
  auto-promotion: every extracted entity and relation waits in the review
  queue (proposed_*, state=pending) and nothing reaches the canonical
  graph without a human decision — the LLM writes proposals, the human writes
  the graph. Covered by a differential end-to-end check
  (npm run test:strict-mode, no LLM required).
• Seeded relationship-type catalog. Migration 20260611000048 seeds
  relation_types with the eight canonical predicates the extraction prompt
  teaches (idempotent; tolerant of ASSIGNED_TO-style naming from existing
  deployments). The dynamic-schema known-predicate filter now has a real
  baseline instead of an empty catalog.
• Dynamic schema (phase 1) — propose and surface. The extraction worker now
  records entity kinds and relationship predicates it observes that aren't in
  the entity_kinds / relation_types catalogs as pending
  schema_proposals rows (provenance-tagged, deduped per workspace by the
  unique type+name constraint; min confidence via
  BRAIN_SCHEMA_PROPOSAL_MIN_CONFIDENCE, default 0.6, inclusive). brain_stats
  gains a schema section and the summary surfaces "Brain proposed N new
  types from your data (pending review)". Promotion stays manual — nothing
  auto-applies: entities with a not-yet-cataloged kind stay in the review queue
  and are not auto-promoted into the canonical graph (so a novel "product"
  can never be mislabeled as an organization or fuzzy-merged into one).
  Migration 20260611000047 (idempotent) extends the proposal_type CHECK
  with entity_kind; the proposal write runs under a savepoint, so a database
  that hasn't applied the migration degrades to "no proposals recorded" instead
  of failing extraction. The extraction prompt now lets the model name a novel
  kind when none of the canonical four fits; the prompt's own example
  predicates are treated as canonical and never proposed as "new".
• Direction regression test. src/extraction-direction.ts defines a
  gold fixture of clearly-directed facts (from examples/demo-corpus, with
  passive-voice traps) and a scoreDirection scorer (deterministic unit tests
  in extraction-direction.test.ts). test/extraction-direction-check.mjs
  (npm run test:direction) measures a model's directed accuracy and fails
  below BRAIN_DIRECTION_MIN_ACCURACY (default 0.8); it skips when no real
  extraction model is configured, so the keyless CI quickstart is unaffected.
• Recommended-model + measured-accuracy note in docs/relationship-extraction.md.
• Local (keyless) vector search via Ollama nomic-embed-text (768d),
  selected with BRAIN_EMBED_PROVIDER=ollama (auto-selected when an Ollama
  base URL is configured and no OpenAI key is present). Runs alongside the
  existing OpenAI text-embedding-3-small (1536d) path; each provider stores
  vectors in its own table (chunks_ollama_nomic / chunks_openai3small), so
  switching providers never causes a dimension clash. brain_search,
  brain_context_pack, and brain_recall_memory join the active provider's
  table, and brain_stats counts embedded chunks across both. Migration
  20260611000046_chunks_ollama_nomic.sql (idempotent) adds the 768d table.
  Semantic search now works with no hosted API key at all.