refactor: AiGateway with adapter pattern (LLM + Codex)

[Pigbibi]

Summary

Extends the existing CodexAuditBridge service into a general-purpose AI Gateway.

Architecture

                    AiGateway
                        │
            ┌───────────┼───────────┐
            ▼           ▼           ▼
    LlmAdapter   CodexAdapter   FutureAdapter
    (API call)   (codex exec)   (your backend)

Changes

• AiAdapter (ABC): base protocol for all AI backends — just implement run()
• LlmAdapter: direct Anthropic/OpenAI API calls (sync response, no repo)
• CodexAdapter: existing codex exec logic (async job, needs repo clone)
• Adapter Registry: _ADAPTER_REGISTRY dict — adding a new backend = one new class
• API keys moved from GitHub repos to VPS env vars

API

• task: "analyze" → LlmAdapter (sync, returns inline)
• task: "execute" → CodexAdapter (async, returns job_id)

Benefits

• API keys in one place (VPS), not duplicated across N repos
• QuantStrategyLifecycle now needs only CODEX_AUDIT_SERVICE_URL
• Adding a new AI backend = new adapter class, no service plumbing changes

Co-Authored-By: Claude noreply@anthropic.com

[github-actions]

🤖 Codex PR Review

⚠️ Review skipped: The Codex review could not be completed.

Codex service request failed: 401 {"status": "error", "error": "OIDC workflow_ref is not allowed"}

Please ensure a human reviewer checks this PR before merging.