v0.5.1

The ChatGPT sign-in's models are the ones it accepts.

Changed

• Rho.Models.catalog(:openai_codex) is Codex's own list for the
  account (GET /backend-api/codex/models, in Codex's order, the first
  its default), fetched and kept by the new Rho.Models.Codex per
  sign-in (per data dir: a loop's :data_dir and :credential, else the
  node's): in memory (an ETS table; Rho.Supervisor starts its owner) and
  in the dir's codex_models.json, fetched again in the background once 6
  hours old (5 minutes after a failed try, one fetch at a time). Before the
  first fetch it is the list shipped in priv/models/openai_codex.json,
  now Codex's of 2026-10-19 (gpt-5.5 and Daybreak Blue, which the
  backend refuses for a ChatGPT account, are gone). A picker should ask
  each time it opens rather than keep the list. catalog/2 takes the
  sign-in (:data_dir, :credential).
• A session on a ChatGPT model the sign-in no longer accepts moves to
  the sign-in's default for good (a model_change entry and a one-line
  notice) when a run starts on it, or when the backend refuses it; the
  refused call is made again on the default, once per run.
• An HTTP error req_llm reports as a bare "HTTP 400" carries the
  provider's own message from the body (Codex's detail).

Added

• Rho.Models.resolve/2: the model a saved spec runs as (a ChatGPT
  model the sign-in no longer accepts resolves to its default).
• config :rho, :codex_models_refresh (default true): false turns
  the background fetch off.

Removed

• Rho.Images.models/1: Rho.Models.Codex.fetch/1 returns the same list
  with names and context windows.