Add Codex Azure OpenAI and Microsoft Foundry MPS test coverage

[masonc08]

What

Adds test coverage for routing Codex through an azure_openai Model Provider Service, backed by main.ucode.ci_azure_openai_mps on the CI workspace (allowlist: gpt-5-nano, gpt-5.6-luna reasoning-capable; gpt-4.1-mini non-reasoning).

e2e — TestAzureOpenAiMps (direct Responses-API probes, no agent binary; runs in the gateway shard, independent of #685):

• reasoning-capable models route with reasoning.effort, and a session can switch between them
• a non-reasoning allow-listed model (gpt-4.1-mini) rejects reasoning.effort with 400 — the parity invariant: a model that wouldn't work with Codex natively must not be papered over by the gateway
• the same model serves a plain (no-reasoning) request
• an off-allowlist model (gpt-4o) is refused with 403

integration — test_ug_configure_codex_azure_openai_mps: the real ug CLI selects the Azure MPS in the configure picker and completes a file task on gpt-5-nano. Requires #685 (adds azure_openai to Codex's provider-type allowlist); until #685 lands, the codex full lane is red.

Plus --codex-azure-provider / --codex-azure-provider-model in run_integration.py and matching conftest fixtures.

Why

Fills the Codex coverage gaps: no live azure_openai coverage, no reasoning-model behavior, and thin OpenAI-MPS coverage generally.

Not covered (deliberately)

Bare ug codex --provider <mps> (no --model) can default to a non-working model — #621 pins the first catalog model, and the MPS discovery call (/ai-gateway/codex/v1/models) carries no per-model reasoning capability, so ucode can't pick a reasoning-capable default. Known limitation, not addressed here.

Test

ruff + py_compile clean; integration contract checks pass. Direct-probe behavior verified live against the MPS (reasoning models 200, gpt-4.1-mini + reasoning.effort 400, gpt-4o 403).

Microsoft Foundry

Foundry speaks the same OpenAI dialect as Azure OpenAI, so test_ug_configure_codex_foundry_mps reuses the Azure CUJ through a shared _run_codex_openai_dialect_mps_cuj helper — the only difference is which MPS the picker selects. Adds --codex-foundry-provider / --codex-foundry-provider-model and matching fixtures. Requires a CI Foundry MPS (flag default main.ucode.ci_foundry_mps) and #685 (adds microsoft_foundry to Codex's allowlist); until the MPS exists the foundry lane errors, as the suite has no capability skips.

This pull request and its description were written by Isaac.