[REGRESSION] Codex: ug codex --provider <MPS> hard-fails at launch when codex/v1/models is disabled, instead of using the supplied --model

[ali-kero]

On ug codex --provider <model-provider-service>, ug now calls GET /ai-gateway/codex/v1/models at launch to fetch the MPS model catalog and raises fatally if it fails. On workspaces where that route isn't enabled, Codex never starts — even though the user passed --model. This worked on an earlier build, so it's a regression.

Environment

• Broken: ug 0.1.0+129.gda2f31d, codex 0.153.4, Databricks CLI 1.14.1, macOS
• Last known working: ucode 0.1.0+101.ge7553f4
• Workspace: fe-sandbox-outcasts-code
• MPS: OPENAI-typed, fronting Azure OpenAI v1; deployment gpt-5.6-sol

Worked before → broken now

• ✅ 0.1.0+101: ug codex --provider <mps> --model gpt-5.6-sol -- "say pong" → pong
• 🚫 0.1.0+129.gda2f31d: same command fails at launch:

ERROR Could not discover Codex models for <mps>: HTTP 404 Not Found:
{"error_code":"ENDPOINT_NOT_FOUND","message":"codex/v1/models is not enabled for this workspace."}

Expected

When codex/v1/models is unavailable, ug should fall back to the user-supplied --model (prior behavior) with a warning, not abort the launch.

Related

• [AIGTWY-4565] Use the MPS model catalog in Codex #549 (AIGTWY-4565) introduced the MPS catalog fetch.
• AIGTWY-4247 — /codex/v1/models currently lists only system.ai.* models, not MPS/external-provider deployments; relevant to whether discovery is even useful for MPS yet.
• codex/v1/models is a per-workspace, off-by-default gateway route (enablement precedents: SUP-35474, SUP-36261).