context-graph: embed messages, entities and edges inside Memgraph for recall

[antejavor]

Closes #416. Implements the decisions in #393; part of map #390 (#415 → #416 → #417 → #418).

What

sessions-graph

• New embeddings.py:
  • embed_session embeds three units through MAGE's embeddings.text, 64 texts per call:
    • messages (Action.text, written by actions-graph since context-graph: message actions carry their plain text #421);
    • entities mentioned in the session's chunks;
    • extracted edges (r.text), reached through those entities because edge properties carry no index.
  • Model: BAAI/bge-small-en-v1.5 by default, which is what the benchmark measured. Each vector records embedding_model. A vector from another model counts as missing and is replaced, so vectors from two models never mix.
  • Errors: EmbeddingUnavailableError covers a missing module (plain Memgraph) or a model that can't load.
• SessionsGraph:
  • embed_session() records embedding_status: completed with embedding_model, or failed with embedding_error.
  • get_pending_embedding_sessions() returns sessions whose embedding failed, never ran, or used another model.
  • Reconciliation, both single and batch, embeds what it wrote. An embedding failure never fails the reconciliation.
• CLI: sessions-graph embed (--session ID | --pending) [--limit N] [--model M]. The model comes from the flag, then recall.embedding_model in the config, then the default.
• Connector:
  • SESSION_END always spawns a detached embed --session, independent of auto_reconcile. It costs no LLM, and the hook never waits on the model.
  • The embed child gets the Memgraph config but not the LLM keys (_child_env(llm=False)).

agent-context-graph

• Config: a recall.embedding_model key, preserved across bootstrap rewrites like auto_reconcile.
• doctor --connector sessions-graph: adds an embeddings check that loads the model inside Memgraph and reports its dimension, or explains that MAGE (2 GiB or more) is needed.
• Bootstrap and docs: bootstrap's hint and the plugin READMEs and skills suggest memgraph/memgraph-mage.

Tests

• Suites:

  Suite	Result
  sessions-graph	78 passed, 1 skipped (OpenAI)
  agent-context-graph	115 passed
  eval	280 passed, 5 skipped

• New end-to-end tests on real MAGE:

  • all three units embedded and tool calls skipped;
  • a second call embeds nothing;
  • a vector from another model is replaced;
  • a model that can't load is recorded on the session;
  • the CLI's success and failure paths.

  The chunk, entity and edge are hand-made in extraction's shape. Reconciliation's own embed call runs only in the existing OpenAI-gated end-to-end test.

• Checked by hand:

  • On plain memgraph/memgraph:3.10.0, embedding raises "There is no procedure named 'embeddings.text'" and the session records failed.
  • doctor's check reports BAAI/bge-small-en-v1.5 (384 dimensions) on MAGE, and fails clearly for a model that can't load.