Sub 2: 5-axis observability primitive + REM run/stage state model

Parent Epic

#12065 — Sub 2 of 9. Parallel to Sub 1.

Premise

Per Discussion #12062 §2.6: REM pipeline state is measurable across 5 distinct axes (Chroma summary count / graph SESSION nodes / per-session ENTITY-RELATION counts / topology-conflict counts / Chroma graphDigested:true flag) which CAN AND DO diverge silently. Currently only axis A is exposed via MCP. GPT live V-B-A 2026-05-27 ~01:19Z showed 76× axis-A-vs-axis-B divergence (1,069 Chroma summaries vs only 14 graph SESSION nodes).

Per Discussion #12062 §2.11: current code uses single graphDigested boolean as completion gate — insufficient since TopologyInferenceEngine returns void (silent failure invisible to flag).

Prescription

Two coupled deliverables:

Part A — 5-axis observability primitive:

• ChromaManager helpers: getUndigestedSessionCount(), getGraphDigestedCount()
• GraphService helpers: getSessionNodeCount(), getSessionEntityCount(sessionId)
• TopologyInferenceEngine: getTopologyConflictCount() (grep sandman_handoff.md for source-session entries)
• MCP exposure: extend manage_database OR add new get_rem_pipeline_state tool
• Per-cycle telemetry log emitting all 5 axis counts

Part B — REM run/stage state model:

Per Discussion #12062 §2.11 schema:

{
  runId, reason, startedAt, completedAt, outcome,
  failurePhase, failureReason,
  perSessionStates: [{sessionId, payloadSizeTokens, memorySessionIngest, triVector, topology, gapSession, graphDigestedFlag, failureReasons}],
  lastSuccessfulPhase, cycleScopePhases
}

Storage: JSONL at .neo-ai-data/rem-runs/<runId>.jsonl (append-only durable; queryable post-restart).

Acceptance Criteria

• AC1: 5 axis-count helpers shipped on ChromaManager + GraphService + TopologyInferenceEngine
• AC2: MCP exposure of axis counts (get_rem_pipeline_state tool OR extension of existing tool)
• AC3: REM run/stage state model JSONL writer integrated into processUndigestedSessions (or into Sub 3's unified method, whichever lands first)
• AC4: Per-phase state tracking includes topology outcome (closes the silent-failure surface of hypothesis #10)
• AC5: Documentation at learn/agentos/rem-state-model.md describing schema + retention policy + consumer patterns
• AC6: Live re-measurement of 5 axes against current production substrate (snapshot before Sub 3 lands)

Avoided Traps

• ❌ Conflate Chroma flag with "session processed" — operator-explicit challenge per Discussion §2.6 update
• ❌ Replace JSONL with graph-nodes-only — graph mutations during REM run risk active-control-plane collateral damage per AC9
• ❌ Skip Part B because hot-fix #12063 helped — state model is independent of cap-raise; still load-bearing for ongoing observability

Related

• Epic #12065
• Discussion #12062 §2.6 (5-axis divergence) + §2.11 (state model schema)
• GPT STEP_BACK empirical anchor: 1069/14/132 axis-divergence at 2026-05-27 ~01:19Z