ai-guardian v0.1.0

Governed observability + governance for on-endpoint local LLM (Ollama) usage — 18 MCP tools. Model allow/deny policy + provenance digest-drift, a deterministic offline prompt scanner (secrets / PII with Luhn / source-code / jailbreak, weighted risk band, redacted findings), and route-through guards (guarded_generate / observe_chat) that scan + policy-gate + record before calling the model. The complement to IGEL AI Armor. Vendored governance harness; remove_model is high-risk with dry-run + double-confirm. Preview — v0.1 passive inventory + opt-in route-through (transparent proxy is v0.2).