Chunk REM topology inference by memory-turn boundary

Context

A live REM run pushed the local graph model beyond its usable context window while processing Memory Core REM graph work. Fresh V-B-A showed this was not a missing primitive: the turn-aligned sessionChunker from #13513 already exists, and #12073 remains the broader Tri-Vector chunk-map/reduce authority.

This ticket exists as the narrow PR close target for the newly-proven topology side path, so the fix does not incorrectly auto-close #12073.

Release classification: boardless P0 runtime hardening; the fix is already scoped to the active REM over-cap burn path.

Live latest-open sweep: checked latest 20 open issues on 2026-06-24; no equivalent narrow topology/chunker bypass ticket found. Recent A2A sweep: only active overlapping claim was @neo-gpt's #12073 lane claim.

The Problem

DreamService.processUndigestedSessions() now fetches complete raw memory documents for a session, but it joins them into one session.document. TopologyInferenceEngine.extractTopology() then receives that joined full-session document and calls the graph provider directly.

That bypasses the turn-boundary chunker, the ConsumerFriction guardrail, structured output bounding, and no-think graph generation settings. A long session can therefore spend minutes in topology/ticket mining and grow the model's generated-token counter far beyond the intended safe band.

The Architectural Reality

• ai/services/graph/sessionChunker.mjs is the deterministic whole-turn chunking primitive from #13513.
• ai/daemons/orchestrator/services/DreamService.mjs has access to rawMemories.documents before it joins them.
• ai/services/graph/TopologyInferenceEngine.mjs is a REM graph consumer separate from SemanticGraphExtractor, and currently uses a full-session prompt path.
• ADR 0019 applies: topology inference must read resolved AiConfig leaves at the use site and fail loud when required leaves are stale or malformed.

The Fix

Thread the complete raw memory document array from DreamService into TopologyInferenceEngine, then make topology inference:

• chunk on whole memory-turn boundaries via sessionChunker;
• subtract the topology prompt envelope from the chunk budget before invoking the model;
• route every chunk through invokeWithGuardrail;
• use graph-provider structured output with the existing no-think config leaf;
• bound topology conflicts to the existing compact handoff cap;
• emit chunk counts into REM phase/per-session state for diagnostics.

Contract Ledger Matrix

Decision Record impact

Aligned with ADR 0019. No ADR change required.

Acceptance Criteria

• DreamService preserves complete raw memory documents for topology inference.
• TopologyInferenceEngine chunks topology input by memory-turn boundary when safe-band pressure requires it.
• Topology provider calls use the ConsumerFriction guardrail.
• Topology output is schema-bounded and capped to the existing compact conflict limit.
• Required AiConfig leaves fail loud when missing or malformed.
• REM topology state includes chunk diagnostic counts.
• Focused unit coverage proves the raw-turn handoff and chunked topology provider path.

Out of Scope

• Full Tri-Vector map/reduce integration remains under #12073.
• No arbitrary string splitting.
• No new cloud/deployment environment variable.
• No change to the public MCP surface.

Avoided Traps

• Do not close #12073 from this PR; that ticket still owns the broader Tri-Vector map/reduce contract.
• Do not create a provider-specific LMS-only fix; the topology path must work for both supported graph providers.
• Do not hide missing config with optional chaining or fallback math.

Related

#12073 #13513 #13624 #13835

Retrieval Hint: query_raw_memories("REM topology full-session burn sessionChunker turnDocuments TopologyInferenceEngine #12073").