dev: add MacBook Pro M5 Pro integration lab for M.I.O.

[01rabbit]

Summary

Adds a reproducible MacBook Pro M5 Pro / Apple Silicon development lab for validating the new M.I.O. cognitive plane together with Azazel-Knowledge and Azazel-Deception before Raspberry Pi HIL.

The intended sequence is:

Mac logical/cross-product validation -> Azazel-Edge #373 adversarial gate -> Raspberry Pi Edge HIL

Added

dev/macbook-m5pro/preflight.sh

• verifies macOS / arm64
• requires Python 3.11+ and Xcode Command Line Tools
• reports Docker, Ollama, GitHub-auth, memory and disk readiness
• installs nothing and changes no network settings

dev/macbook-m5pro/bootstrap.sh

• clones/updates:
  • 01rabbit/Azazel-Edge -> feat/mio-cognitive-shadow-core (PR #380)
  • 01rabbit/Azazel-Knowledge -> main
  • 01rabbit/Azazel-Deception -> main
• creates a separate .venv per product so different Fabric pins do not collide
• provisions isolated Knowledge development state outside all Git repos
• redacts the one-time Knowledge admin token from bootstrap output and does not persist it
• mints and retains only a scoped ingest,query,read Edge-development token
• generates a random AZ-06 shadow HMAC secret
• stores runtime secrets only in ~/azazel-m5pro-lab/runtime.env with mode 0600
• runs the focused M.I.O. shadow/replay test set
• optionally runs full Knowledge/Deception suites with AZAZEL_FULL_TESTS=1
• optionally pulls qwen3.5:2b and qwen3.5:0.8b with AZAZEL_PULL_MODELS=1

dev/macbook-m5pro/start-shadow-lab.sh

• starts Azazel-Knowledge API on 127.0.0.1:8070
• starts the Azazel-Knowledge worker offline, without --online, so local event/flow/reaction/deception-observation spools and behavior processing work without scheduled internet feed pulls
• starts AZ-06's existing shadow/replay service on 127.0.0.1:8071
• uses the scoped Edge ID + HMAC transport key
• keeps AZ-06 live_execution=disabled
• verifies Ollama at 127.0.0.1:11434 without managing or exposing it
• prints the exact M.I.O. local-model replay command

dev/macbook-m5pro/README.md

• defines the Mac-first validation gate
• documents the model-comparison workflow (0.8B / 2B / later larger reference model)
• separates logical/cross-product validation from Raspberry Pi Linux/network/HIL validation
• documents the Knowledge and Deception authority boundaries

Safety / authority

This lab does not:

• run the Azazel-Edge installer
• install or enable systemd units
• alter macOS packet filtering/routing
• invoke nftables, iptables or tc
• connect M.I.O. directly to Action Arbiter enforcement
• enable live AZ-06 engagement
• expose Knowledge/Deception on non-loopback interfaces by default
• use a public/cloud LLM fallback

M.I.O. remains advisory-only and the current Edge branch continues to produce executable=false recommendations.

Development gate before Raspberry Pi

The Mac phase should prove:

• M.I.O. reasoning/replay stability
• Knowledge absent/slow/malformed/wrong degradation
• Deception unavailable/malformed/stale degradation
• cross-trace and prompt/delimiter/Unicode injection rejection
• reproducible closed-loop replay
• deterministic Arbiter remains the only authority
• Azazel-Edge #373 findings are resolved or explicitly accepted

Only after that should Azazel-Edge move to Raspberry Pi HIL for Raspberry Pi OS, systemd, real NICs, nftables/tc, resource pressure, thermals and long-duration behavior.

Related work

• 01rabbit/Azazel-Edge#380 — governed M.I.O. cognitive shadow core
• 01rabbit/Azazel-Edge#360 — M.I.O. Offline White-Hacker Reasoning Core epic
• 01rabbit/Azazel-Edge#373 — adversarial architecture/security review gate
• 01rabbit/Azazel#62 — Azazel System terminology/state-model unification