[codex] Pin multi-agent system per thread

[aibrahim-oai]

Why

A thread and its descendants must keep one multi-agent communication system for their lifetime. Re-resolving from flags during resume, fork, or child spawning can make related agents incompatible.

What changed

• Resolve the existing rule once: MultiAgentV2 -> v2, otherwise Collab -> v1, otherwise omitted.
• Store the resolved value in session state and turn context.
• Restore persisted values on resume and inherit them through forks, children, and model switches.
• Pin legacy rollouts on first resume while preserving rollout timestamps.
• Use the resolved turn-context value for tool exposure, hints, notifications, interruption markers, lifecycle behavior, and concurrency gates.
• Keep existing numeric tuning, depth behavior, and review suppression unchanged.

Verification

Added focused integration coverage for flag fallback, legacy rollout pinning, resume and fork behavior, child inheritance, model switches, v1 and v2 tools, hints, notifications, lifecycle behavior, and v2 concurrency gates. Ran just fmt. Validation is running in online CI.

Stack

1. Metadata storage: [codex] Persist multi-agent version metadata #25153
2. This PR: [codex] Pin multi-agent system per thread #25168
3. ModelInfo overlay: [codex] Add model multi-agent system overlay #25155

Supersedes #25032.