chatgpt-new: sign-in ends after an hour because every refresh returns invalid_grant #1422

Symptom. A chatgpt-new sign-in stops working about an hour after graff login chatgpt-new. The renewal is refused, and once the access token expires every request fails with invalid_token until you sign in again.

Error. OpenAI's token endpoint answers every grant_type=refresh_token request with:

HTTP/1.1 400 Bad Request

{"error": "invalid_grant"}

graff keeps that answer in the sign-in record as last_refresh_error.

Root cause. Upstream. graff sends the documented request (refreshed in src/oauth_chatgpt.zig on feat/chat-gpt-new): the issued client_id, the latest refresh_token and resource, never before earliest_refresh_at. OpenAI refuses it even for a never-used token inside the renewal window. T3 Code, pi and nolune send the same four fields.

Fix. Nothing to change in graff until OpenAI answers. Reported upstream with a reproduction: docs/upstream/openai-sign-in-refresh-invalid-grant.md. Workaround: run graff login chatgpt-new again.