Cloud lite workers: self-hosted model provider and operator-driven jobs #9

Cloud lite workers can only use the subscription login today: the runtime writes forced_login_method="chatgpt" into every Codex configuration, requires an auth broker, and clones with a GitHub token. Teams that serve an open-weight model themselves (for example with vLLM) cannot point the workers at it.

Proposed, all optional and off by default:

1. plan.modelProvider: a self-hosted OpenAI-compatible (Responses API) provider. Codex gets a config.toml with the provider, the model's real context window and compaction point; the key reaches only the Codex process (fleet.deploy(plan, { providerKey })), not agent commands, journals or traces.
2. Read-only deploy-key repository access (fleet.deploy(plan, { deployKey })) with GitHub's published host key pinned, instead of a GitHub token.
3. plan.prepare: a hook that runs once per machine after the pinned clone and returns each agent's folder, writable roots and instructions.
4. plan.supervision: 'operator' (agents start idle; every turn is an operator message), plan.sandbox (workspace-write / read-only), and plan.boardTools: false.
5. Less trace volume: streaming deltas are not recorded (plan.traceDeltas: true restores them), and plan.monitorIntervalMs sets how often machines read operator commands.

This has run on real fleets, including a 190-agent coding run on one B300 serving Nemotron 3.5 Super VL. Ship it with focused offline tests and a short guide in PUBLIC_SDK.md, then release 0.4.0.