v3.1.0

Added

• Adaptive memory layer, extended across TypeScript, Rust, and Python:
  • Per-context activation (NodeActivation.context): reinforceNode/
    reinforce_node gain an optional context argument that reinforces an
    independently-decaying, additional lens alongside the global score.
    getContextActivation/get_context_activation reads it back.
  • Inhibition (NodeActivation.inhibition/lastInhibitedAt): a new
    suppressNode/suppress_node primitive with its own decay half-life,
    subtracted from score only at the final read/ranking layer so a
    suppressed node stays re-evaluable rather than permanently invisible.
  • Budgeted, diversity-aware ActivationEngine.workingMemory/
    working_memory: an options form (tokenBudget, costOf,
    diversityLambda, similarityOf) implementing a memory-flavoured
    maximal-marginal-relevance selection, backward compatible with the
    existing (limit, minScore) call shape.
  • Memory classes (episodic/semantic/procedural/entity): a
    per-node memoryClass override falling back to a schema-level
    NodeTypeDefinition.memoryClass default, driving differentiated
    score/importance decay half-lives via ActivationConfig.classHalfLives.
  • Confidence and provenance fields on PolyNode: confidence, source,
    observedAt, derivedFrom, supersedes, contradicts — ordinary,
    optional node fields (not activation state), settable via addNode,
    updateNode, and patchNode (which now also supports these as
    top-level set/unset/increment/compareAndSet paths, not only
    data.*, in all three languages).
  • supersede(id, supersededId, amount?, reason?): a contradiction
    primitive that records supersedes, adds a SUPERSEDED_BY edge, and
    suppresses the superseded node without deleting it.
  • consolidate(node, sourceIds, options?): a consolidation primitive
    that writes a higher-level node with derivedFrom merged from its
    sources, adds CONSOLIDATED_FROM edges, and suppresses the sources.
  • ActivationEngine.recordFeedback(id, wasUseful, learningRate?):
    learned scoring weights — nudges pulse's composite
    semantic/graph/recency/usage weights from outcome feedback instead of
    requiring hand-tuning.
  • Fixed a latent double-decay bug in the TypeScript
    ActivationEngine.effective, found while adding class-resolved
    half-lives: it re-decayed the already-decayed result of
    getActivationState a second time, silently breaking any non-default
    half-life under real elapsed time.
  • See specification/data-model.md sections 1.5–1.10 and
    docs/API.md for full details.

Fixed

• Restored an edge-id integrity check in the Rust graph's rebuild_edge_index
  (rejecting empty ids) that had been dropped while adding support for
  independent edge IDs.
• Python DirectoryStorage now implements real sync/sync_dir (fsync),
  bridged from the native storage adapter, so its advertised fsync
  capability is honest.
• Rust attribute_ranges queries now resolve nested dot-path fields the same
  way attributes does, instead of only matching flat top-level fields.
• Rust's in-memory GraphQuery index selection now excludes a sparse index
  when the queried value is null, matching PersistedGraphQuery.
• Rust count_nodes now derives its maxNodesVisited candidate set the same
  way query_nodes does, instead of always falling back to the full store
  size.
• The TypeScript BinaryStoreAdapter now dedups retried batches by
  transactionId as well as operationId, matching the Rust/Python
  idempotency contract.
• Rust's SyncServer::submit adds a clientId:seq fallback dedup (evicted
  alongside the ops ring buffer), matching the TypeScript server.
• Python's SyncServer.receive stops evaluating conflict() for an
  operation authorize() already rejected, matching the TypeScript server.