Fix Sandman graph provider routing

Context

Operator reported a P0 regression while running npm run ai:run-sandman manually on May 26, 2026. The run initialized Memory Core, found undigested sessions, successfully ran deterministic Memory/Session graph ingestion for the first session, then failed during Tri-Vector extraction:

buildGraphProvider: unsupported modelProvider 'gemini'. Expected one of: 'ollama', 'openAiCompatible'.

The operator also observed that the resulting handoff file is incomplete and stated the Sandman/Dream pipeline should never use Gemini for this path.

The Problem

The closed provider-routing work in #11965 deliberately made graph dispatch support ollama and openAiCompatible, and providerDispatch.mjs correctly rejects gemini rather than silently falling back. However, Sandman/Dream graph call sites still read the generic Memory Core modelProvider value. The Tier-1 default remains gemini, so manual Sandman can crash before completing semantic extraction and Golden Path handoff generation.

This is not a request to add Gemini support to graph dispatch. The regression is that the Sandman/Dream graph-generation path is using the wrong provider axis.

The Architectural Reality

• ai/config.mjs:70 and ai/config.mjs:79 default chatProvider / modelProvider to gemini.
• ai/services/graph/providerDispatch.mjs:19 documents graph dispatch support as ollama or openAiCompatible; providerDispatch.mjs:51-55 throws for unsupported providers.
• ai/services/graph/SemanticGraphExtractor.mjs:96-100 passes aiConfig.modelProvider into buildGraphProvider().
• ai/services/graph/TopologyInferenceEngine.mjs:57-61 does the same for topology extraction.
• ai/services/graph/GoldenPathSynthesizer.mjs also instantiates graph provider dispatch from aiConfig.modelProvider.
• learn/agentos/DreamPipeline.md:321 already documents the Dream extraction provider default as openAiCompatible, so current code and docs disagree.
• Related closed authority: #11965 AC included "Dream/REM graph-mutator provider reachability is not hardwired to the wrong provider family." PR #11966 implemented provider dispatch and tests that gemini is unsupported for graph dispatch.

The Fix

Introduce or reuse a graph-generation provider selector that defaults Sandman/Dream graph work to openAiCompatible, while preserving explicit ollama support and explicit failure for unknown provider names.

The fix should update all graph-generation call sites consistently:

• SemanticGraphExtractor.executeTriVectorExtraction()
• SemanticGraphExtractor.extractMessageConcepts()
• TopologyInferenceEngine.extractTopology()
• GoldenPathSynthesizer.synthesizeGoldenPath()

Manual Sandman should probe and use the same provider axis it will later dispatch through. It must not claim OpenAI-compatible readiness while the graph dispatch path still points at gemini.

Contract Ledger Matrix

Decision Record impact

Aligned with ADR 0015 graph-store backend posture and with #11965 provider-routing closure. No ADR amendment expected.

Acceptance Criteria

• npm run ai:run-sandman no longer routes graph extraction through gemini when no graph-specific provider is configured.
• Graph-generation call sites share one selector/helper instead of each reading generic modelProvider directly.
• Explicit ollama graph routing remains supported.
• Unsupported graph provider values fail with a clear diagnostic before partial REM/handoff work proceeds.
• Tests cover the regression: modelProvider: 'gemini' with no graph override routes graph work to openAiCompatible, not to gemini.
• PR evidence includes the operator reproducer, focused unit tests, and a syntax/static check. Full local ai:run-sandman may be skipped if it would mutate live Memory Core, but the reason must be documented.

Out of Scope

• Adding Gemini support to providerDispatch.mjs.
• Changing embedding provider defaults.
• Reworking all Tier-1 provider key migration.
• Running destructive or broad Sandman cycles against production Memory Core without explicit operator direction.

Related

• Parent provider-routing work: #11965 / PR #11966
• Parent config epic: #10103
• Dream pipeline docs: learn/agentos/DreamPipeline.md

Origin Session ID: d84ad2f1-c71c-4ce4-bec7-1167b9183637

Handoff Retrieval Hint: query_raw_memories("runSandman unsupported modelProvider gemini DreamService SemanticGraphExtractor providerDispatch")