feat: implement auth, rate limiting, and cost attribution from AgentRelay reference spec

[stackbilt-admin]

Context

The AgentRelay architecture spec has been committed to docs/reference-agentrelay-architecture.md as both a design reference and scaffold quality benchmark (from project archaeology, aegis#319).

This 901-line architecture package covers the exact features the MCP gateway needs but doesn't yet have.

Features to Implement (from spec)

1. Scoped API Key Auth

• API keys as primary auth for the data plane
• Per-key permission scopes (tool-level access control)
• Key lifecycle: create, rotate, revoke with audit trail
• Hash-based key storage (SHA-256, never store plaintext)

Reference: Already extracted and proven in smart_revenue_recovery_adf/src/shared/tenant-auth.ts — can port directly.

2. Per-Key Rate Limiting

• Sliding window rate limiting per API key
• 429 response with retry-after header
• Configurable limits per key/tier

3. Usage Tracking + Cost Attribution

• Per-request metering (tool, key, tokens, latency)
• Cost rollup per API key, per tenant
• Dashboard visibility into usage patterns
• Ties into existing Stripe billing

4. Sync + Async Tool Calls

• Synchronous path for immediate results
• Async path: queue job → return job ID → poll for status
• Streaming responses for long-running tools

Reference Files

• docs/reference-agentrelay-architecture.md — full architecture spec (Product → UX → Risk → Architect → TDD → Sprint)
• docs/reference-agentrelay-architecture.json — structured JSON version

Priority

Auth and rate limiting are prerequisites for public MCP gateway access. Cost attribution feeds into the billing system. Async tool calls are a nice-to-have for v1.