Write turn links and edge source sentences at extraction time

[antejavor]

Part of #390

Question

Hybrid's index step derives, after extraction and inside the eval, what extraction could have written:

• entity-to-turn links;
• each edge's source turn (exact, via valid_at);
• each edge's source sentence.

Should the GLiNER2 backend write them at ingest instead? And in what shape: MENTIONED_IN to the turn's Action rather than the session Chunk, or beside it? r.turn / r.text on edges? Segment would need to carry the source node's id.

Depends on

The lift ticket. Not gated any more: the map judges the system end to end (#391). Decisions are in the comments.

[antejavor]

Decisions (grilling, 2026-10-01)

1. Provenance anchor. The chunk stays the session. Segment gains an optional, opaque source_id, which sessions-graph fills with the turn's action_id. unstructured2graph never learns what an Action is. Rejected: chunk = turn, which would split chunk-scoped entities per turn.
2. Mention → turn. A sources list on the existing (entity)-[:MENTIONED_IN]->(:Chunk), merged by union, so re-ingesting is idempotent. Rejected: one parallel relationship per turn; a second write pass in sessions-graph.
3. Edge identity. One edge per turn, keyed on (type, head, tail, chunk, source_id). Each statement keeps its own valid_at, which also fixes the last-write-wins SET r.valid_at. The collapse key in aingest_chunk gains source_id. Without a source_id, behaviour is today's.
4. r.text. The sentence or sentences covering both endpoint spans inside the segment, capped at about 300 characters. The full turn is one hop away through r.source_id.
5. Scope. Provenance plus r.role (user or assistant) from the segment. r.role tells the two speakers apart; it isn't a filter. Assistant content shapes what the user does, and questions ask about it. Extraction quality goes to its own ticket.
6. Old graphs. link_turns and MENTIONED_IN_TURN are deleted, with no fallback. The eval graph is rebuilt once with provenance, together with the per-question user fix, and the baseline is re-measured on it.

Smaller defaults: when identical turns are merged into one segment, it takes the first turn's action_id and valid_at. All of this comes from the spans GLiNER2 already returns, so there are no extra model calls.

Tests, against a real Memgraph with the GLiNER2 fakes:

• a two-turn chunk gets exact sources on its mention links;
• each edge gets source_id, text and role;
• the same fact in two turns becomes two edges;
• re-ingesting changes nothing.