Support a Codex-native mode that can use the current Codex session as the LLM runner

### Problem Statement

First of all, thank you for open-sourcing Open Code Review and for adding Codex skill/plugin support. The project is very useful, especially the deterministic diff handling and structured review workflow.

I would like to use Open Code Review from inside Codex via the installed skill/plugin. Today, the Codex integration invokes the local `ocr` CLI, and `ocr` still requires a separately configured LLM endpoint/API key. For users who already have a Codex subscription or token allowance, this means the Codex integration cannot use their existing Codex quota and requires another provider setup or extra API billing.

This is a bit confusing from a user perspective because the workflow is launched inside Codex, but the actual model calls are still managed outside Codex.

### Proposed Solution

It would be great to have a Codex-native mode for the Codex integration.

Possible shape:

- A documented mode such as `ocr review --audience agent --llm-runner codex`, or a dedicated Codex skill/plugin path.
- In this mode, Open Code Review could keep the deterministic parts of the pipeline, such as diff/file selection, grouping, rules, and output formatting.
- The LLM judgment/review step would be delegated to the current Codex session, so users can use their existing Codex quota instead of configuring `OCR_LLM_URL` / `OCR_LLM_TOKEN`.

If direct access to the Codex model runner is not technically possible, another helpful option might be to expose an `ocr` command that only prepares the deterministic review context as JSON/Markdown. The Codex skill could then ask Codex itself to perform the review from that context.

### Alternatives Considered

Current workarounds I found:

- Configure a separate OpenAI/Anthropic/OpenAI-compatible endpoint for `ocr`.
- Use a local model through Ollama/LM Studio.
- Use Codex's native code review without Open Code Review.

These work, but they either require extra provider setup/cost, or they lose the Open Code Review pipeline that makes this project valuable.

### Affected Area

Review Agent / LLM interaction

### Additional Context

Related docs/PRs I looked at:

- The README says the Codex plugin registers a skill that runs `ocr review --audience agent`.
- PR #69 notes that the Codex integration does not make Codex the internal LLM runner for OCR.

I completely understand there may be product/API limitations here. Even a short documentation note explaining the limitation and recommended Codex-native workaround would be very helpful. Thanks again for the great project!

Start with the README's Codex skill documentation and the existing `ocr review --audience agent` entry point, then read PR #69 for the current integration boundary. Determine whether a Codex-native runner is technically available; done means either a supported runner/context-preparation path or a clear documented limitation and workaround.

Written by the indexing model from the issue text.