Inference Gateway

Implement gateway logic in this order:

1. Validate and normalize request input.
2. Estimate prompt + generation token budget before admission.
3. Apply tenant token bucket and global concurrency checks.
4. Compute KV-cache pressure and shed early when threshold is exceeded.
5. Build scheduler job with immutable metadata (tenant_id, adapter_id, request_id, token budget).
6. Emit gateway metrics for accepted and rejected traffic.

Operational Rules

• Prefer deterministic rejection at ingress over best-effort processing.
• Return 429 Too Many Requests for rate-limit or pressure rejection.
• Return 503 Service Unavailable for internal scheduler unavailable states.
• Preserve idempotent request identifiers for trace stitching.
• Keep adapter selection pure (no side effects in resolver).

Contracts

• Read request-contract.md before changing API schema.