v1.6.0 — Credibility Hardening

What's New

Production Features Wired into Server

• API key authentication (X-API-Key header, configurable via api_keys)
• Correlation IDs (X-Request-ID echo/generate on every response)
• Audit logging wired into /v1/review and /v1/process endpoints
• Tenant routing via X-Tenant-ID header + /v1/tenants management API
• Rate limiting via slowapi (rate_limit_rpm config)
• Streaming oversight via WebSocket /v1/stream with StreamingKernel

Streaming Debug Mode

• streaming_debug=True on StreamingKernel emits per-token snapshots (window_avg, trend_drop, accumulated_tokens)
• TokenEvent.debug_info and StreamSession.debug_log for full trace

E2E Benchmark Proof

• --baseline mode: unguarded hallucination rate
• --compare mode: side-by-side delta table (baseline vs guarded)
• --output-json for CI artifact upload
• New benchmarks/ragbench_eval.py for real-world RAG pipeline evaluation
• Regression suite: 6th test test_e2e_heuristic_delta (catch=70%, FPR=10%)

Documentation

• Threshold Tuning Guide — heuristic vs NLI, domain presets, sweep instructions
• KB Ingestion Tutorial — GroundTruthStore, VectorGroundTruthStore, ChromaDB, multi-tenant
• Production Checklist — 12-item deployment readiness checklist
• Streaming Architecture — ASCII diagram, NLI construction, debug mode docs

Test Coverage

• 835 tests pass, 0 failures across Python 3.10/3.11/3.12
• 19 new tests (auth, audit, tenant, streaming debug)