v0.12.1

Fixes the knowledge-graph calls against the real server contract (checked live against
ghcr.io/dakera-ai/dakera:0.12.0; the v0.11.108 server has the same request and response
types). Before this release none of them matched what the server sends or needs.

Changed (breaking for calls that never worked)

• memoryLink(sourceId, targetId, { agentId, label? }): POST /v1/memories/{id}/links
  requires agent_id (the agent owning both memories); the old
  memoryLink(sourceId, targetId, edgeType) sent none, so every call failed with
  422 INVALID_REQUEST (missing field agent_id). The server records every explicit link as
  linked_by, so the edge-type argument is gone; passing a string as the third argument (the old
  call) or omitting agentId throws a TypeError before any request. label is sent, but the
  server does not store it.
• DakeraDelegationHelper.linkDelegation() (TealTiger integration) takes agentId and creates a
  linked_by edge; getDelegationChain() follows linked_by edges, the type the server stores
  delegation links as (it never had a delegated_from type, so the old chain was always empty).

Fixed

• Graph edges (memoryGraph, agentGraphExport, knowledgeQuery, knowledgeExport, the
  link answer): the server sends from_id, to_id, edge_type, weight, created_at and no
  edge id. Edges are now normalized: from_id / to_id plus the older source_id /
  target_id (same values) and id: '', so code reading either name works. GraphEdge gains
  from_id / to_id.
• memoryLink() answer: the server answers {from_id, to_id, edge_type}; GraphLinkResponse
  now has those fields, and edge is built from them (weight 1.0, as the server stores explicit
  links; created_at 0 because the answer omits it). It used to be undefined.
• memoryGraph(): the server answers {root_id, depth, node_count, nodes} with each node's
  edges on the node (nodes[].edges), no top-level edges and no content_preview /
  importance. MemoryGraph gains node_count; GraphNode gains edges (the two never-sent
  fields are now optional); edges collects every node's edges (it was undefined). The types
  filter is applied client-side and no longer sent (the route has no such parameter).
• memoryPath() sent ?target=; the route reads ?to=, so every call failed with
  400 INVALID_REQUEST (missing field to). The answer is {from_id, to_id, path, hop_count};
  GraphPath gains those fields and keeps source_id / target_id / hops as the same values
  (the never-sent edges field is removed).
• agentGraphExport(): the route always answers JSON
  {agent_id, namespace, node_count, edge_count, edges} and ignores format, which is no longer
  sent (the argument stays so existing calls compile; use knowledgeExport(agentId, 'graphml')
  for GraphML). GraphExport now has namespace and edges instead of the never-sent
  format / data.
• memoryEntities(): the server answers {entities, count} without memory_id;
  memory_id is now filled from the requested id and count is typed.
• recall() / searchMemories() accept tags (the server's request has it; it could not be
  sent).
• Path segments of the graph routes are URL-encoded.

Tests

• Unit tests use the server's real shapes (the old fixtures encoded the wrong contract), and
  check the to parameter, the link body, the TypeError for old calls, client-side type
  filtering, tags, and the agent_id query of getMemory / updateMemory.
• src/integration.test.ts has a live round trip against the CI server: store two memories,
  update, link with agentId, graph, path, export, knowledge query and entities.

What's Changed

Bug Fixes

• fix: graph, link and entities calls match the server (0.12.1) @ferhimedamine (#248)

Full Changelog: v0.12.0...v0.12.1