feat(infra): add local Mac profile and privacy validation demos

[entrkjm]

Feature Proposal: Local Infrastructure for Privacy & Sovereignty

The "Why": Beyond Cost Savings

Hi ROMA Team! I've been testing ROMA on my M4 MacBook.
While saving API costs is beneficial, the real value of local execution is Data Sovereignty.
Many developers work in environments where:

1. Air-Gapped / Offline: No internet access allowed.
2. Privacy-Critical: Sending code to cloud APIs (OpenAI/Anthropic) is a security violation.
3. Censorship-Free: Need for unaligned/uncensored local models.

I created a "Local Mac Infrastructure Kit" to enable these workflows.

PR Contents

This PR provides the infrastructure layer to make local execution viable:

1. local_mac.yaml: Optimized for Apple Silicon (Extended timeouts 600s, lightweight config).
2. Friendly Error Patch: Modifies runtime.py to catch ConnectionRefused errors and guide users to start Ollama.
3. Validation Demos: Scripts to prove the privacy and utility of the setup.

Validation & Analysis

I tested this extensively on an M4 Max with Llama 3.1 8B. Here is the honest breakdown:

The Good: Privacy & Code Security

• Task: examples/security_audit_demo.py (Scanning code for secrets).
• Result: PASSED. The agent successfully identified hardcoded credentials.
• Value: Zero data egress. Perfect for auditing sensitive repositories.

The Challenge: Reasoning Capabilities

• Task: examples/doc_slave_demo.py (Complex JSON schema for documentation).
• Result: LIMITATION FOUND. Llama 3.1 8B struggled with complex Pydantic schemas (Outputted int instead of string IDs).
• Takeaway: Local execution works for Retrieval/Scanning tasks, but complex Reasoning/Planning may require larger models (70B) or specific prompt tuning.

Conclusion

This PR doesn't claim to replace GPT-4.
Instead, it opens the door for private, offline, and sovereign AI development on ROMA.
It builds the bridge so the community can start tuning local agents for their specific needs.