[Feature]: Pass configured reasoning_effort through openai-codex Responses path

[Oldrich333]

Issue Origin

Observed or reproduced in a real environment

Problem Statement

For vlm.provider = openai-codex, OpenViking cannot honour a configured reasoning effort today.

Two gaps:

1. VLMConfig (openviking_cli/utils/config/vlm_config.py) has no reasoning_effort field (extra = forbid), and _build_vlm_config_dict() does not pass one into the VLM instance config.
2. Even when OpenAIVLM puts reasoning_effort on the Chat Completions kwargs, CodexCompletionsAdapter._create_response (openviking/models/vlm/backends/codex_responses_adapter.py) builds a Responses API request with only model / instructions / input / store / tools and drops reasoning_effort.

Net effect for subscription Codex summarization: the Responses call always uses the model default effort. Operators cannot set high / xhigh / max from ov.conf.

Related but distinct from #4066 (LiteLLM per-call thinking translation). This issue is specifically the openai-codex Responses adapter + config surface for reasoning.effort.

Proposed Solution

1. Add optional reasoning_effort: str | None to VLMConfig and pass it through _build_vlm_config_dict() / credential builders.
2. In CodexCompletionsAdapter._create_response, when kwargs contain a truthy reasoning_effort, set:

   response_kwargs["reasoning"] = {"effort": reasoning_effort}

   When unset/None, omit the field so the model default remains.
3. Document vlm.reasoning_effort in ov.conf examples for openai-codex.

Alternatives Considered

• Leave effort at model default (works for many summarization workloads; this is what we run today).
• Out-of-tree bind-mount overlays (works, but forks behaviour from upstream and must be rebased on every image bump).
• Static extra_request_body — does not reach the openai-codex Responses adapter today either.

Feature Area

Model Integration

Use Case

Production knowledge stores that summarize resources via openai-codex (ChatGPT/Codex subscription) need the same effort control agent runtimes already expose (low…xhigh). Without a first-class config + Responses mapping, operators cannot raise effort for hard corpora or lower it for cheap bulk summarization.

Example API (Optional)

{
  "vlm": {
    "provider": "openai-codex",
    "model": "gpt-5.6-luna",
    "reasoning_effort": "xhigh"
  }
}

When reasoning_effort is omitted, behaviour stays as today (model default).

Additional Context

Happy to send a small PR if maintainers want this as a code change rather than design-only. Platform / Model Integration seems closest; please re-route if needed.