Run manual Pi live validation on gpt-5.6-luna at maximum reasoning

[clkao]

End value

Manual Pi live validation uses the same OpenAI model and reasoning level as the Codex pull-request lane: gpt-5.6-luna with model_reasoning_effort="max".

Current behavior

• Pi live tests are manual and are not a pull-request CI job.
• The Pi runner defaults to openrouter/openai/gpt-5.4.
• Direct OpenAI authentication uses openai/gpt-5.4.
• The current Pi launch passes --model but does not set a reasoning effort.

Required change

• Change the Pi live default to the provider-qualified gpt-5.6-luna model for each supported authentication path.
• Set model_reasoning_effort="max" through Pi's supported provider or session configuration mechanism.
• Apply the setting to the First Officer and dispatched Pi ensigns.
• Do not encode reasoning effort as prompt prose.
• Update the manual run instructions and examples in docs/runtime-live-ci.md.

Acceptance criteria

• A manual Pi live run records gpt-5.6-luna for both the root and child runs.
• Durable run evidence records maximum reasoning configuration for both runs.
• A negative control fails if either run falls back to gpt-5.4 or omits maximum reasoning.
• Existing Pi front-door, dispatch, and durable-report checks pass.
• The implementation documents the exact supported Pi configuration surface used for reasoning effort.

Scope boundary

This issue changes the existing manual Pi live path. It does not add Pi to pull-request CI or change the Sonnet, Opus, or Codex lanes.