[awf] Disable rubber-duck sub-agent when Copilot is the engine

[lpcox]

Problem

When running agentic workflows with the Copilot engine, the system prompt includes instructions to proactively invoke the rubber-duck sub-agent for plan validation and critique. This adds significant token overhead and latency for every non-trivial task — launching a separate Haiku-model agent, waiting for its response, then processing the feedback.

For Copilot engine runs, this is wasteful because:

1. The rubber-duck agent uses additional model calls that count against the workflow's token budget
2. It adds latency (sync sub-agent calls block the main agent)
3. Copilot's built-in reasoning is generally sufficient for the tasks these workflows handle
4. The defense-in-depth value is lower when the primary model already has strong reasoning capabilities

Proposed Solution

When the workflow engine is copilot, the compiled lock file should either:

• Omit the rubber-duck collaboration instructions from the system prompt entirely, or
• Include a directive like "Do not use the rubber-duck agent" in the system prompt

This could be implemented as:

• A per-engine default in the compiler (rubber-duck enabled for claude/codex, disabled for copilot)
• A workflow-level opt-in/opt-out field (e.g. rubber-duck: false)
• Both (engine default + explicit override)

Context

The rubber-duck instructions are injected via the Copilot CLI system prompt and encourage the agent to call task(agent_type: "rubber-duck") before implementing non-trivial changes. While valuable for some engines/tasks, it's pure overhead for Copilot engine workflows that are already well-scoped.