Enhance agent teams/subagents orchestration: config-driven, hot-reload, no hardcoded NL routing

[chumyin]

Background

Current agent delegation needed deeper separation and intelligence between synchronous agent teams and asynchronous subagents, with runtime configurability and natural-language operability.

Goals

• Separate orchestration semantics for team delegation vs background subagents.
• Ensure natural-language driven updates are configuration-driven (no hardcoded keyword triggers for specific toggles/counts).
• Support hot apply for enable/disable and count limits without restart.
• Keep defaults enabled.

Proposed scope

• Add explicit runtime config sections:
  • [agent.teams]: enabled, auto_activate, max_agents
  • [agent.subagents]: enabled, auto_activate, max_concurrent
• Extend delegate agent profile metadata for non-hardcoded selection:
  • enabled, capabilities, priority
• Introduce shared agent selection strategy used by both delegate and subagent_spawn.
• Add model_routing_config action set_orchestration for NL-driven config mutation.
• Keep tools registered and enforce toggles at execution-time so config updates hot-apply.
• Add tests for runtime toggle/concurrency hot apply and metadata persistence.

Acceptance criteria

• delegate and subagent_spawn both support agent omitted/auto when corresponding auto-activation is enabled.
• Runtime disable for teams/subagents is enforced immediately from config.
• Subagent concurrent limit is runtime-configured (not fixed constant).
• Selection favors metadata/capability/priority rather than hardcoded agent names.
• Tests cover new behavior and pass.

[chumyin]

Tracking implementation PR: #2420