fix(agent): reset api_mode to chat_completions when switching to MoA (#54259)

[Sahil-SS9]

Fixes #54259

Description

When a session is switched to a MoA preset (via /moa or a persisted MoA-model session), switch_model() left agent.api_mode unchanged from whatever the previous provider used. If the previous provider was a codex slot (api_mode = "codex_responses"), the conversation loop's primary/acting call was routed through client.responses.create() against the non-HTTP placeholder base_url = "moa://local", returning HTTP 404 and falling back to a reference model — the final answer was never the aggregator's.

The init path in agent_init.py:724 already forces agent.api_mode = "chat_completions" when the provider is MoA. This change applies the same force inside the runtime switch_model() path so a session that started on a codex slot and is then switched to a MoA preset reaches a consistent state.

Verification

• Targeted tests — pytest tests/run_agent/test_switch_model_moa_api_mode.py -v: 3 passed
  • test_switch_to_moa_from_codex_resets_api_mode — repro: codex_responses → moa must end in chat_completions
  • test_switch_to_moa_from_chat_completions_keeps_api_mode — regression: chat_completions → moa stays chat_completions
  • test_switch_to_moa_invalidates_transport_cache — transport cache is cleared on the switch
• S1 (Secrets) — clean
• S2 (Personal refs) — clean
• C1 (Conventional commit) — fix(agent): ...
• F1 (Focused) — only agent/agent_runtime_helpers.py and a new test file
• Compile check — ast.parse passes for both files

[teknium1]

Already on main — switch_model pins agent.api_mode = "chat_completions" in the provider == "moa" branch as of #55579, matching agent_init.py. Same fix landed in parallel. Closing #54259 is covered.