Codex-native harness timeout when routed through OpenAI-compatible gateway (OpenRouter).

## Summary

When a `codex-native` worker is routed through an OpenAI-compatible **gateway** provider (here: OpenRouter, `kind: key`, `wire_api: chat`), the Codex app-server never starts a thread. `wait_for_thread_started` times out after 30s, the bridge thread state is never written, and the next executor turn fails with the misleading error **`Codex native bridge state is missing`**.

The same gateway works fine for the `pi` harness, and the Codex CLI works fine on its own ChatGPT subscription, so this is specific to `codex-native` over a gateway provider.

## Environment

- omnigent 0.1.0 (uv tool install)
- codex-cli 0.139.0 (`codex login status` → "Logged in using ChatGPT")
- Python 3.12, macOS (arm64)
- Agent: bundled `examples/polly` `codex` sub-agent (`harness: codex-native`, no explicit model)

## Reproduction

`~/.omnigent/config.yaml` providers block (secrets elided) makes a gateway the default for the `openai` family:

```yaml
providers:
codex:
cli: codex
kind: subscription # exists, but NOT a default for openai
openrouter:
kind: key
default: [openai, pi] # gateway is the openai-family default -> captures codex-native
openai:
base_url: openrouter.ai/api/v1
wire_api: chat
models: { default: moonshotai/kimi-k2.6 }
```

Dispatch any `codex-native` task (e.g. a polly `codex` sub-agent). It is routed to the gateway because `default_provider_for_harness("codex-native")` resolves the `openai` family, whose o...