feat: auth gate, rate limiting, and cost attribution

[stackbilt-admin]

Summary

Implements the three security-critical features from issue #18 (AgentRelay reference spec):

• Rate limiting — Sliding window per-tenant rate limiter using RATELIMIT_KV. Tier-based limits: free=20/min, hobby=60, pro=300, enterprise=1000. Returns 429 with standard Retry-After and X-RateLimit-* headers on all responses.

• Cost attribution — Per-tool credit costs with quality multipliers for image_generate. Reserves quota via edge-auth consumeQuota RPC before each tool call, then commits or refunds based on outcome. Zero-cost tools (read-only lookups) skip quota enforcement entirely.

• Scope enforcement — Mutation tools (LOCAL_MUTATION, EXTERNAL_MUTATION, DESTRUCTIVE) require the generate scope. tools/list filters the catalog to only show tools the session has access to. API keys with read-only scopes cannot invoke mutations.

New files

• src/rate-limiter.ts — Fixed-window rate limiter with KV TTL for auto-cleanup
• src/cost-attribution.ts — Tool cost registry, quota reservation/settlement
• test/rate-limiter.test.ts — 8 tests
• test/cost-attribution.test.ts — 17 tests

Modified files

• src/gateway.ts — Wire rate limiting after auth, quota reserve before tool call, scope check on tools/call
• src/types.ts — Add checkQuota, consumeQuota, commitOrRefundQuota to AuthServiceRpc; add RATELIMIT_KV to GatewayEnv
• wrangler.toml — Add RATELIMIT_KV namespace binding
• All test files — Updated mocks to include quota methods and RATELIMIT_KV

Test plan

• 25 new tests pass (rate-limiter: 8, cost-attribution: 17)
• All 116 non-oauth tests pass (0 regressions)
• TypeScript compiles cleanly (tsc --noEmit)
• Pre-existing oauth-handler.test.ts failures are unchanged (6 tests, same on main)
• Integration test: deploy to staging, send 21 requests with free-tier key, verify 429 on 21st
• Integration test: verify X-RateLimit-Remaining header decrements
• Integration test: call image_generate with ultra_plus quality, verify 40-credit cost

Closes #18

🤖 Generated with Claude Code