docs(copilot): clarify subagent execution model and orchestration patterns

[digitarald]

Summary

Addresses user confusion about subagent behavior based on internal Slack feedback. Clarifies how runSubagent works and when to use it.

Key changes

Execution model

• Subagents are synchronous (blocking) — main agent waits for results
• Multiple subagents can run in parallel when spawned together
• Subagents inherit model + tools but NOT instructions or conversation history
• Added mermaid diagram showing the flow

UI visibility

• Subagents appear as collapsible tool calls
• Collapsed: shows agent name + current tool
• Expanded: shows full prompt, tool calls, and result

Invocation patterns

• Clarified that subagents are typically agent-initiated, not user-invoked
• Added guidance on defining subagent behavior in custom agent instructions

Custom agent integration

• agents array overrides disable-model-invocation: true for explicit allowlisting
• Custom agents can specify their own model, tools, instructions

Orchestration patterns (new section)

• Coordinator and worker pattern: Feature Builder with Planner, Plan Architect, Implementer, Reviewer
• Multi-perspective code review: parallel review subagents with prompt-based focus areas
• Worker agents use user-invokable: false and tailored tool access

Style compliance

• Removed hypotheticals (would/could)
• Added periods to numbered steps
• Used active voice throughout