Sub 1: Silent-failure root-cause investigation across 13 hypotheses

Parent Epic

#12065 — Sub 1 of 9. MUST land FIRST per GPT migration blast-radius sweep ordering.

Premise

REM pipeline silently produces no fans-glow despite the orchestrator's dream task code calling processUndigestedSessions(). Discussion #12062 §2.4 + §2.4.1 enumerate 13 distinct silent-failure hypotheses. Until we identify which hypothesis is actively firing in production, the unified orchestrator-as-SSOT shape (Sub 3) could replicate the failure at a different layer.

Prescription

V-B-A each of the 13 hypotheses against production state. For each, document: (a) is it actively firing? (b) what evidence proves/disproves it? (c) what mitigation is needed?

13 hypotheses (full descriptions in Discussion #12062 §2.4):

1. DreamService.isProcessing stuck flag
2. Provider probe gate skipped for gemini/ollama
3. Cadence silent-loop (markStarted + markFailed without work)
4. MaintenanceBackpressureService indefinite-deferral
5. Lease held by orphan
6. DreamService.initAsync hung on dependency
7. autoDream config flip
8. Pre-PR #11966 OpenAiCompatible-hardwire (last-2-days only)
9. invokeWithGuardrail size-cap pre-check (PRIMARY — addressed by hot-fix #12063/#12064)
10. TopologyInferenceEngine.extractTopology returns void (most insidious)
11. 3-retry JSON parse exhaustion
12. Provenance edge culling
13. Lazy edge queue overflow

Acceptance Criteria

• AC1: V-B-A report per hypothesis (active/inactive/inconclusive + evidence + mitigation strategy)
• AC2: Forensic protocol documented for each hypothesis (e.g., "how to detect isProcessing-stuck: check .neo-ai-data/wake-daemon/* for orphan lease + DreamService log debug-line freq")
• AC3: Inventory document at learn/agentos/sandman-silent-failure-forensics.md (or similar canonical location) serves as runbook for future regressions
• AC4: Hypothesis-mitigation specs handed off to Sub 3 (unified REM method) as preconditions for implementation

Avoided Traps

• ❌ Fix-the-symptom-without-V-B-A — must actually verify each hypothesis empirically; don't guess
• ❌ Skip hypothesis 9 since #12063 handles it — STILL document the forensic protocol so future cap regressions are diagnosable
• ❌ Defer mitigation specs to Sub 3 implementation — Sub 3 needs the specs INPUT; Sub 1 produces them OUTPUT

Related

• Epic #12065
• Discussion #12062 §2.4 + §2.4.1 (hypothesis enumeration)
• Memory anchor f04dd0ba (the deferral pattern this sub structurally prevents)