fix(proxy): resolve provider-prefixed models in Codex Desktop

[Filip-Stastny-Philips]

Summary

Codex Desktop could show an available GPT model but send the conversation to a different model. The proxy did not recognize deployment names with a provider prefix, such as openai.gpt-5.6-luna, so it treated the selected model as unavailable and used a fallback. In the reported case, selecting GPT-5.6 Luna routed to GitHub Copilot GPT-5.3 Codex and failed.

Changes

Make Desktop model selection work with provider-prefixed catalogs, using the same naming convention already supported by the VS Code connector. This completes the Desktop routing support missed by #603.

Impact

Selecting gpt-5.6-luna now resolves to the available openai.gpt-5.6-luna deployment. The fix also covers other provider namespaces and GPT-6/6.1 models.

Validation

• Passed 48 local routing checks using the staging configuration: all six primary GPT models and their aliases. Also checked regional fallback mappings.
• Passed license check, lint, typecheck, build, and related proxy tests run by the commit hook.
• Windows end-to-end conversation confirmation is pending.
• Local secrets scan skipped with approval because no container engine is available; GitHub CI must perform it.

Checklist

• Self-reviewed
• Manual testing performed (local routing; Windows conversation pending)
• Documentation updated (local investigation report)
• No breaking changes