v0.14.0

Changed — behavior

• Resolver source-precedence taxonomy expanded 7 → 9 tiers (#328). A new
  agent-observed:<model>:<session-ref> tier is inserted at rank 5 (below
  wikipedia, above claude:), which re-ranks the lower tiers
  (claude: → 6, script: → 7, model-prior: → 8, unsourced → 9). Conflict
  resolutions that compare claims across these tiers can therefore reach a
  different winner than in 0.13.x. The change to the taxonomy itself is
  additive (no field removed); the behavior change is in ranking.
• repair --backfill-sources rewrites the source: scalar of existing
  DEFAULTED claude:inferred memories (#328). With --apply, claims whose
  origin transcript shows the user stated them are lifted to user:<ref>
  (tier 1) and claims derived from in-session artifacts to
  agent-observed:<...> (tier 5). This raises their precedence, so future
  resolutions over pre-existing data can change outcome. Dry-run by default;
  only DEFAULTED inferred claims are touched; idempotent (a confirmed claim
  gets inferred_verified: true and is never re-examined).
• Opinions no longer lose to precedence (#327). Pairs classified
  claim_kind: opinion with different (or unknown) asserters resolve to the
  new attribute_both action — both stay active, neither is superseded.
  Same-asserter dated opinions still supersede (newer wins). This changes the
  outcome of opinion-vs-opinion conflicts that previously picked a precedence
  winner.

Added

• Incremental indexing for both search backends (#348). Whole-file
  content-hash diffing rebuilds only changed/new/deleted pages; an unchanged
  corpus is a sub-second no-op instead of a full re-embed. Adds --full and a
  config seam for the embedding model (default all-MiniLM-L6-v2 unchanged).
• On-demand athenaeum ingest / athenaeum reindex (--incremental|--full) (#349)
  — change-gated compile + index with one-line JSON summaries, single-flight.
• athenaeum session-end for cross-agent same-day recall (#350). A
  change-gated compile-then-index entrypoint: a remember written in one
  session becomes recallable in another after that session ends, without
  waiting for the nightly librarian. No-op when nothing changed.
• claim_kind classification + attribute_both resolver action (#327),
  including asserter-identity comparison (same / different / unknown)
  with a keep-both fallback when identity is unavailable.
• repair --backfill-sources (#328) — re-classify DEFAULTED
  claude:inferred provenance from origin transcripts to user-stated,
  agent-observed, or confirmed-inferred.
• Scoped claims: org/locale dimensions + three-way overlap verdict (#329)
  (DISJOINT / OVERRIDE / OVERLAP), disjoint scopes short-circuit to
  not-a-conflict without an LLM call.
• Temporal recall --as-of <date> view and per-claim compiled validity (#308).
  A historical read-time view over validity windows, and per-source validity
  stamped into compiled entries.
• athenaeum compile --as-of <date> --out <dir> — historical recompiled wiki
  view (#359). Distinct from the read-time recall --as-of filter: re-runs the
  deterministic C3 blend with as_of threaded into the per-member validity
  predicate, writing to a scratch dir (live wiki never touched, no LLM spend), so
  a member expired-as-of-today is re-included when the date precedes its
  valid_until close. Valid-time rewind; transaction-time replay is documented as
  deferred (the frontmatter model lacks per-member assertion timestamps).
• athenaeum serve honors KNOWLEDGE_RAW_PATH / KNOWLEDGE_WIKI_PATH (#355)
  — each env var overrides its root individually; otherwise falls back to
  <path>/raw|wiki. Makes athenaeum's MCP server a drop-in for the standalone
  local knowledge server.

Fixed

• Resolver aggregate eval floor recalibrated + JSON-repair retry (#345).
  Two mislabeled golden fixtures corrected, the golden set enlarged 5 → 8, the
  floor re-derived with slack, and propose_resolution now retries once with a
  strict-JSON reminder when the first response has no parseable JSON object.
• TestBatchSyncEquivalence no longer hangs locally (#362). The batch-mode
  test double now reports an immediately-completing batch at create time, so the
  poll loop's real 30s sleep is skipped under test (suite runs in seconds, not
  ~60s/test); production poll cadence unchanged.