MCP: roll out a change set to a slice of live traffic

An AI assistant connected over the MCP server can now serve a testing change set to a sample of new conversations before it is promoted. The staged variants run against real traffic while the rest of the AI Agent stays on the live version.

What’s new

Start a rollout: edit_agent_behavior accepts set_rollout with changeset_id, percentage (1–100), and max_conversations. The rollout completes the moment the cap is reached.
Stop a rollout: edit_agent_behavior accepts stop_rollout with changeset_id. No further conversations are sampled in. Conversations already pinned to the change set keep serving its variants until they end.
Combined cap: percentages across all active rollouts on the same Agent must sum to at most 100.