BYE-340 route GPT-5.6 labels through Codex runtime

[b-yelverton]

What changed

• Map the terra, luna, sol, and bare gpt-5.6 labels to the requested GPT-5.6 model IDs and select the Codex runner.
• Carry bare or effort:-prefixed low/medium/high labels through runner config, defaulting to medium.
• Serialize the selected effort into the Codex app-server config so it reaches OpenAI as reasoning.effort.
• Fail loudly for recognized GPT-5.6 labels that conflict with a Claude-family selector or an existing non-Codex session.
• Log unknown model-family-shaped labels before applying the normal fallback.

Root cause

RunnerSelectionService only recognized gpt-* model labels. terra/luna/sol therefore selected the default Claude runner, and the existing Codex modelReasoningEffort field was never serialized by AppServerCodexBackend.

Validation

• pnpm install --frozen-lockfile
• pnpm build
• pnpm typecheck
• pnpm test:packages:run — all package tests green; EdgeWorker 66 files / 744 tests and Codex 12 files / 71 tests
• pnpm --filter './apps/*' test:run — CLI 5 files / 94 tests; F1 no-test pass
• targeted Biome check and git diff --check

Linear: BYE-340

[Connoropolous]

this looks great, just ask that the change in default reasoning effort, and documentation on how to use reasoning overrides via labels make it into the changelog entry