v5.4.0 — Security-Hardened LLM & Hermes Recognition

v5.4.0 — Security-Hardened LLM & Hermes Recognition

🔴 Fixed (v5.3.0 audit)

• F-01: Substring bypass — "OWNER" in first_line would match negations
• F-02: Blocking HTTP in decision loop — could freeze state machine
• F-03: LLM could grant presence alone — now prevented by TrustPolicy
• F-04: --llm-api-key leaked in process list
• F-05: Zero tests for 1,010 new lines of security code

🏗️ Architecture (v5.4.0)

• TrustPolicy (T3→T0): LLM asks, deterministic table authorizes
• Strict grammar: 1 token per line, NFKC, no substring fallback
• InferenceWorker: non-blocking, budget + circuit breaker
• Hermes integration: authenticated, 6 audited events
• Egress gate: cloud providers require --llm-accept-egress

📦 Providers

• OpenAI (GPT-4o-mini), Anthropic (Claude 3 Haiku), Ollama (local)
• Hermes Agent (Bearer token, endpoint probe)

✅ 368 tests passing