feat(codex): send version header so backend serves newer models (gpt-5.6-luna)

[skulidropek]

Problem

For Codex (ChatGPT) subscriptions the router forwards POST /v1/responses to
https://chatgpt.com/backend-api/codex/responses, preserving the caller's model verbatim.
The ChatGPT Codex backend gates newer models behind a recent client version, advertised via
a version HTTP header (the Codex CLI sends version: <cli-version>). The router sets
chatgpt-account-id, openai-beta and originator, but not version.

As a result, newer models are rejected even though the account has access:

# model gpt-5.6-luna, identical body/token
without version header ->  404 {"error":{"message":"Model not found gpt-5.6-luna",...}}
with    version: 0.144.1 -> 200 event: response.created ...

Verified end-to-end: the same ChatGPT account that returns Model not found through the router
serves the model fine via the Codex CLI (which sends the version header), and adding the header
to the router's request makes gpt-5.6-luna work while gpt-5.4 / gpt-5.5 keep working.

Change

Send a version header on Codex requests, defaulting to 0.144.1 and overridable via
CODEX_CLIENT_VERSION, mirroring the Codex CLI. Adds a unit test (codex_headers_include_version)
and a changelog fragment.

Notes

The router remains a transparent proxy — this only advertises a client version so the backend
exposes the models the account is entitled to. Model availability is still governed by the account.