A per-IP ceiling on device-flow polling
POST /api/v1/device/token now refuses more than 7,200 polls per hour from a single IP address. A normal `makespdf login` polls one code 12 times a minute, far under the ceiling, so nothing changes for it.
POST /api/v1/device/token — the endpoint makespdf login polls while you approve the code in your browser — now has a ceiling of 7,200 polls per hour per client IP. Over the ceiling it returns the same slow_down error, with the same retry interval, that the existing 12-polls-per-minute per-code limit already returns. Any client that handles slow_down needs no change.
The number is the most the flow can legitimately produce: 60 code requests per hour per IP (the limit on POST /api/v1/device/code) × a 10-minute code lifetime × 12 polls per minute = 7,200. One login polling one code once every five seconds uses a small fraction of that.
The old limit was keyed only on the device_code in the request body, so a caller willing to invent codes could take a fresh bucket — and a fresh Durable Object — per value. The per-code limit is unchanged and still applies on top of the new per-IP one; only the invented-code path is now bounded.