unstructured2graph/sessions-graph: write turn provenance at extraction time

[antejavor]

Closes #392. Part of map #390. The decisions are recorded on #392.

Why

Hybrid retrieval needs to know which turn each entity and fact came from. Until now the GLiNER2 backend dropped that at write time: a session is a single chunk, so entities linked only to the whole session. The eval then guessed the turn afterwards (link_turns in #389), by timestamp or substring match, which links a short or common entity name to every turn that contains the string. The backend already knows the exact turn, because it extracts one window per segment, and a segment is a turn.

What changes

• Segment.source_id: an optional, opaque id that unstructured2graph never interprets. sessions-graph fills it with the turn's action_id. When identical turns are merged into one segment, it takes the first turn's id.
• MENTIONED_IN.sources: the turn ids that mention each entity in the chunk, merged by union, so re-ingesting is idempotent.
• One edge per turn. upsert_extracted_relationships keys on (type, head, tail, chunk, source_id). A fact said in two turns becomes two edges, each with its own valid_at. Before, they collapsed into one edge, and SET r.valid_at kept whichever turn was written last.
• Every edge carries source_id, role (user or assistant) and text: the sentence or sentences covering both endpoint spans, at most 300 characters. The full turn is one hop away through source_id. role tells the two speakers apart; it isn't a filter.
• Without a source_id, e.g. plain document ingestion, keys and behaviour are unchanged.

No extra model calls: all of it comes from spans GLiNER2 already returns.

Tests

All run against a real Memgraph with the GLiNER2 fakes:

• exact sources for an entity mentioned in two turns;
• an edge's turn, speaker and sentence;
• the same fact in two turns becomes two edges with their own times;
• re-ingesting with provenance is idempotent;
• the sentence span and the 300-character cap.

The sessions-graph reconciliation e2e test now checks that the edge's source_id is the user message's action_id, and that sources lists both turns.

Results, run locally:

• unstructured2graph: 142 passed, 10 skipped (they need gliner2 or LightRAG);
• the real-GLiNER2 e2e tests, run separately in a gliner2 environment: 3 passed;
• sessions-graph: 71 passed;
• ruff and ty are clean.

Follow-ups

• context-graph-eval: hybrid retrieval strategy (typed graph as an index into turns) #389: switch the hybrid read path to these fields and delete link_turns and MENTIONED_IN_TURN.
• Rebuild the eval graph once, together with Eval: one user per question, retrieval scoped to it #396 (one user per question), and re-measure the baseline.