- Add new Claude rate-limit helper utilities to detect unified header-based quota rejections (5h/7d), handling case-insensitive headers and missing canonical forms. - Compute deterministic retry duration from `Retry-After` and header reset timestamps, preferring the longest applicable unified window so cooldowns can be applied consistently. - Introduce reusable rate-limit classification logic to distinguish header-authoritative rejections from request-scoped errors (e.g. fast-mode entitlement failures) for correct credential/model cooldown behavior. Closes: #4874