SemanticGraphExtractor: emit typed edges to Memory + Session nodes

Context

Parent sub-epic (#10143) phase 2. After Memory + Session are nodes (phase 1), extractor attaches extracted entities (concepts/classes/methods/tickets) to those nodes via typed edges instead of sessionId-scalar metadata.

The Problem

Extractor today produces entity nodes with only sessionId scalar metadata — no graph edge to the anchoring Session or Memory. Consumers can't traverse from a concept back to its source memory or from a ticket-reference back to the authoring session via the graph.

The Architectural Reality

• ai/daemons/services/SemanticGraphExtractor.mjs runs Gemma 4-31B (via Ollama) with a Zod-validated tri-vector schema.
• Current emission is node-only; provenance encoded (if at all) as metadata properties.
• Post-phase-1, Memory and Session nodes exist as edge targets.

The Fix

• Extend Gemma4 extraction prompt to emit provenance edges alongside entities:
  • MENTIONED_IN(Concept → Memory)
  • DISCUSSED_IN(Class | Method → Session)
  • REFERENCED_BY(Issue → Memory)
• Update Zod schema to validate the new edge shapes
• Post-processing maps emitted edges to graph addEdge calls
• Backward compat: if Memory/Session nodes don't exist yet (pre-migration), edges queue for lazy back-fill rather than drop

Acceptance Criteria

• Gemma4 extraction prompt updated with provenance edge emission examples
• Zod schema validates new edge types; rejections surface in tri-vector-errors telemetry
• Edges land in graph: integrity check confirms extracted entities have provenance edges to Memory/Session (when source is post-phase-1)
• Existing extracted-entity emission unchanged (no regression beyond additive edges)
• Telemetry: REM cycle comparison pre- vs post-shipping shows edge growth proportional to entity count

Out of Scope

• Changing which entities are extracted
• Retroactively attaching edges to historical nodes (via lazy back-fill)
• Human review of extracted edges for correctness (validation via integrity checks + telemetry)

Avoided Traps

• Scalar metadata for provenance instead of edges. Rejected. Metadata strings can't be traversed; edges can.
• Hardcoded edge emission in post-extractor code. Rejected. Gemma4 has per-extraction context to pick which memory sourced which entity; post-extractor would re-derive. Emission-from-LLM is a small prompt change with high semantic fidelity.

Related

• Parent: #10143
• Blocked by: phase 1 (Memory + Session ingestion)
• Files: ai/daemons/services/SemanticGraphExtractor.mjs, extraction prompt templates, Zod schema

Origin Session ID: 71dc3cd8-d39d-48e1-ac62-e240ca67d1a5